topic
stringclasses 1
value | RAG
stringclasses 1
value | prompt
stringlengths 16.2k
20.1k
|
|---|---|---|
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
abstract: Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Large language models memorize parts of their training data. Memorizing short
snippets and facts is required to answer questions about the world and to be fluent
in any language. But models have also been shown to reproduce long verbatim
sequences of memorized text when prompted by a motivated adversary. In this
work, we investigate an intermediate regime of memorization that we call non-
adversarial reproduction, where we quantify the overlap between model responses
and pretraining data when responding to natural and benign prompts. For a variety
of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up
to 15% of the text output by popular conversational language models overlaps with
snippets from the Internet. In worst cases, we find generations where 100% of the
content can be found exactly online. For the same tasks, we find that human-written
text has far less overlap with Internet data. We further study whether prompting
strategies can close this reproduction gap between models and humans. While
appropriate prompting can reduce non-adversarial reproduction on average, we
find that mitigating worst-case reproduction of training data requires stronger
defenses—even for benign interactions.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
abstract: What Makes Large Language Models Reason in (Multi-Turn) Code Generation?
Prompting techniques such as chain-of-thought have established themselves as a
popular vehicle for improving the outputs of large language models (LLMs). For
code generation, however, their exact mechanics and efficacy are under-explored.
We thus investigate the effects of a wide range of prompting strategies with a focus
on automatic re-prompting over multiple turns and computational requirements.
After systematically decomposing reasoning, instruction, and execution feedback
prompts, we conduct an extensive grid search on the competitive programming
benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama
3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that
consistently improve performance across all models with small and large sampling
budgets. We then show how finetuning with such an optimal configuration allows
models to internalize the induced reasoning process and obtain improvements in
performance and scalability for multi-turn code generation.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Lines of Thought in Large Language Models
abstract: Lines of Thought in Large Language Models
Large Language Models achieve next-token prediction by transporting a vector-
ized piece of text (prompt) across an accompanying embedding space under the
action of successive transformer layers. The resulting high-dimensional trajecto-
ries realize different contextualization, or ‘thinking’, steps, and fully determine
the output probability distribution. We aim to characterize the statistical proper-
ties of ensembles of these ‘lines of thought.’ We observe that independent tra-
jectories cluster along a low-dimensional, non-Euclidean manifold, and that their
path can be well approximated by a stochastic equation with few parameters ex-
tracted from data. We find it remarkable that the vast complexity of such large
models can be reduced to a much simpler form, and we reflect on implications.
Code for trajectory generation, visualization, and analysis is available on Github
at https://github.com/rapsar/lines-of-thought.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Research Town: Simulator of Research Community
abstract: Research Town: Simulator of Research Community
Large Language Models (LLMs) have demonstrated remarkable potential in scien-
tific domains, yet a fundamental question remains unanswered: Can we simulate
human research communities using LLMs? Addressing this question could deepen
our understanding of the processes behind research idea generation and inspire
the automatic discovery of novel scientific insights. In this work, we propose
RESEARCHTOWN, a multi-agent framework for simulating research communi-
ties. Within this framework, the real-world research community is simplified and
modeled as an agent-data graph (i.e.community graphs), where researchers and
papers are represented as agent-type and data-type nodes, respectively. We also
introduce TextGNN, a text-based inference framework that models diverse research
activities (e.g., paper reading, paper writing, and review writing) as specific forms
of a generalized message-passing process on the agent-data graph. To evaluate the
quality of research simulation, we present RESEARCHBENCH, a benchmark that
uses a node-masking prediction task for scalable and objective assessment. Our
experiments reveal three key findings: (1) RESEARCHTOWN effectively simulates
collaborative research activities by accurately predicting the attribute of masked
nodes in the graph; (2) the simulation process in RESEARCHTOWN uncovers in-
sights, like not every author contributes equally to the final paper, which is aligned
with real-world research communities; (3) RESEARCHTOWN has the potential to
foster interdisciplinary research by generating reasonable paper ideas that span
across domains.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
abstract: GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
The powerful capabilities of Large Language Models (LLMs) have led to their grow-
ing use in evaluating human-generated content, particularly in evaluating research
ideas within academic settings. Existing solutions primarily rely on prompt-based
LLM methods or fine-tuned lightweight language models for idea evaluation. How-
ever, these methods are often unstable and struggle to comprehend the complex
semantic information embedded in the ideas, impeding their ability to perform high-
quality evaluations. To address the above challenges, we propose GraphEval,
a lightweight graph-based LLM framework for idea evaluation. Our insight is
that a complex idea can be broken down into comprehensible viewpoint-nodes
using small prompted LLMs. These viewpoint-nodes can then be linked together
through edges created from LLM-based relation extraction and/or BERT similarity
scores. The created viewpoint-graph can be used to conveniently propagate scores
across viewpoint-nodes to improve the robustness of the idea evaluations. In par-
ticular, we propose two lightweight graph-based methods for idea evaluation: (1)
GraphEval-LP: a training-free label propagation algorithm that propagates quality
labels from known viewpoint-nodes to unknown nodes; (2) GraphEval-GNN: a
Graph Neural Network (GNN) that is trained to predict the quality labels given
the observed graph with minimal computation resources. Moreover, to overcome
LLM’s limitation in objectively assessing the novelty of ideas, we further propose
a novelty detection model to GraphEval-GNN to enhance its capability in judging
idea novelty. Experiments on two datasets show GraphEval improves F1 scores
by at least 14% with low computation and API costs. Additionally, GraphEval
can effectively detect plagiarized ideas. Our codes for GraphEval is released at
https://github.com/ulab-uiuc/GraphEval.
title: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
abstract: Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Recent advancements in large language models (LLMs) have highlighted the im-
portance of extending context lengths for handling complex tasks. While tradi-
tional methods for training on long contexts often use filtered long documents,
these approaches lead to domain imbalances, limiting model performance. To
address this, techniques like random document concatenation (Standard) and
similarity-based methods (KNN, ICLM) have been developed. However, they
either sacrifice semantic coherence or diversity. To balance both aspects, we in-
troduce Quest, a query-centric data synthesis method aggregating semantically
relevant yet diverse documents. Quest uses a generative model to predict po-
tential queries for each document, grouping documents with similar queries and
keywords. Extensive experiments demonstrate Quest’s superior performance on
long-context tasks, achieving remarkable results with context lengths of up to 1M
tokens and confirming its scalability across various model sizes.
Figure 1: The Needle-in-a-Haystack task evaluates a model’s ability to retrieve specific information
(the needle) from a large collection of documents (the haystack). Following LongVA (Zhang et al.,
2024a) and LWM (Liu et al., 2024), where the x-axis represents the document length and the y-axis
indicates the position of the ”needle” within the document, ranging from 25K to 1M tokens. To the
best of our knowledge, Quest is the first base model (without instruction tuning) to achieve 100%
accuracy with a 1M context length.
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
abstract: Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
As the context limits of Large Language Models (LLMs) increase, the range of
possible applications and downstream functions broadens.
In many real-world
tasks, decisions depend on details scattered across collections of often disparate
documents containing mostly irrelevant information. Long-context LLMs appear
well-suited to this form of complex information retrieval and reasoning, which has
traditionally proven costly and time-consuming. However, although the develop-
ment of longer context models has seen rapid gains in recent years, our under-
standing of how effectively LLMs use their context has not kept pace. To address
this, we conduct a set of retrieval experiments designed to evaluate the capabilities
of 17 leading LLMs, such as their ability to follow threads of information through
the context window. Strikingly, we find that many models are remarkably thread-
safe: capable of simultaneously following multiple threads without significant loss
in performance. Still, for many models, we find the effective context limit is sig-
nificantly shorter than the supported context length, with accuracy decreasing as
the context window grows. Our study also highlights the important point that to-
ken counts from different tokenizers should not be directly compared—they often
correspond to substantially different numbers of written characters. We release
our code and long context experimental data.
title: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
abstract: Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Humor is previously regarded as a gift exclusive to humans for the following
reasons. Humor is a culturally nuanced aspect of human language, presenting
challenges for its understanding and generation. Humor generation necessitates
a multi-hop reasoning process, with each hop founded on proper rationales. Al-
though many studies, such as those related to GPT-o1, focus on logical reasoning
with reflection and correction, they still fall short in humor generation. Due to the
sparsity of the knowledge graph in creative thinking, it is arduous to achieve multi-
hop reasoning. Consequently, in this paper, we propose a more robust frame-
work for addressing the humor reasoning task, named LoL. LoL aims to inject
external information to mitigate the sparsity of the knowledge graph, thereby en-
abling multi-hop reasoning. In the first stage of LoL, we put forward an automatic
instruction-evolution method to incorporate the deeper and broader thinking pro-
cesses underlying humor. Judgment-oriented instructions are devised to enhance
the model’s judgment capability, dynamically supplementing and updating the
sparse knowledge graph. Subsequently, through reinforcement learning, the rea-
soning logic for each online-generated response is extracted using GPT-4o. In this
process, external knowledge is re-introduced to aid the model in logical reasoning
and the learning of human preferences. Finally, experimental results indicate that
the combination of these two processes can enhance both the model’s judgment
ability and its generative capacity. These findings deepen our comprehension of
the creative capabilities of large language models (LLMs) and offer approaches to
boost LLMs’ creative abilities for cross-domain innovative applications.
title: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
abstract: Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Recent advancements in large language models (LLMs) have sparked optimism
about their potential to accelerate scientific discovery, with a growing number of
works proposing research agents that autonomously generate and validate new
ideas. Despite this, no evaluations have shown that LLM systems can take the
very first step of producing novel, expert-level ideas, let alone perform the entire
research process. We address this by establishing an experimental design that
evaluates research idea generation while controlling for confounders and performs
the first comparison between expert NLP researchers and an LLM ideation agent.
By recruiting over 100 NLP researchers to write novel ideas and blind reviews of
both LLM and human ideas, we obtain the first statistically significant conclusion
on current LLM capabilities for research ideation: we find LLM-generated ideas
are judged as more novel (p < 0.05) than human expert ideas while being judged
slightly weaker on feasibility. Studying our agent baselines closely, we identify
open problems in building and evaluating research agents, including failures of
LLM self-evaluation and their lack of diversity in generation. 1
title: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
abstract: Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
A wide range of LM applications require generating text that conforms to
syntactic or semantic constraints. Imposing such constraints can be naturally
framed as probabilistic conditioning, but exact generation from the resulting
distribution—which can differ substantially from the LM’s base distribution—is
generally intractable. In this work, we develop an architecture for controlled LM
generation based on sequential Monte Carlo (SMC). This SMC framework allows
us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information
during the course of generation. By comparing to a number of alternatives and
ablations on four challenging domains—Python code generation for data science,
text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with
little overhead, our approach allows small open-source language models to outper-
form models over 8× larger, as well as closed-source, fine-tuned ones. In support
of the probabilistic perspective, we show that these performance improvements
are driven by better approximation to the posterior distribution. Our system builds
on the framework of Lew et al. (2023) and integrates with its language model
probabilistic programming language, giving users a simple, programmable way
to apply SMC to a broad variety of controlled generation problems.
https://github.com/probcomp/genparse
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
ai_researcher | True | You are an expert researcher. Now I want you to help me brainstorm some new research project proposals on the topic of: using large language models to generate novel research ideas.
Here are some relevant papers on this topic just for your background knowledge:
title: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
abstract: Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
Large language models (LLMs) have demonstrated remarkable in-context learning
(ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits
two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised
function learning tasks where prompts are constructed with i.i.d. input-label pairs.
This i.i.d. assumption diverges significantly from real language learning scenarios
where prompt tokens are interdependent. (b) Lack of Emergence Explanation.
Most literature answers what ICL does from an implicit optimization perspective
but falls short in elucidating how ICL emerges and the impact of pre-training
phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm,
auto-regressive next-token prediction (AR-NTP), which closely aligns with the
actual training of language models. Specifically, within AR-NTP, we emphasize
prompt token-dependency, which involves predicting each subsequent token based
on the preceding sequence. To address (b), we formalize a systematic pre-training
and ICL framework, highlighting the layer-wise structure of sequences and topics,
alongside a two-level expectation. In conclusion, we present data-dependent, topic-
dependent and optimization-dependent PAC-Bayesian generalization bounds for
pre-trained LLMs, investigating that ICL emerges from the generalization of
sequences and topics. Our theory is supported by experiments on numerical linear
dynamic systems, synthetic GINC and real-world language datasets.
title: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
abstract: Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Large Language Models (LLMs) struggle with long-context reasoning, not only
due to the quadratic scaling of computational complexity with sequence length but
also because of the scarcity and expense of annotating long-context data. There
has been barely any open-source work that systematically ablates long-context
data, nor is there any openly available instruction tuning dataset with contexts sur-
passing 100K tokens. To bridge this gap, we introduce a novel post-training syn-
thetic data generation strategy designed to efficiently extend the context window
of LLMs while preserving their general task performance. Our approach scalably
extends to arbitrarily long context lengths, unconstrained by the length of avail-
able real-world data, which effectively addresses the scarcity of raw long-context
data. Through a step-by-step rotary position embedding (RoPE) scaling training
strategy, we demonstrate that our model, with a context length of up to 1M tokens,
performs well on the RULER benchmark and InfiniteBench and maintains robust
performance on general language tasks.
title: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
abstract: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Scientific discovery contributes largely to human society’s prosperity, and recent
progress shows that LLMs could potentially catalyze this process. However, it
is still unclear whether LLMs can discover novel and valid hypotheses in chem-
istry. In this work, we investigate this central research question: Can LLMs au-
tomatically discover novel and valid chemistry research hypotheses given only a
chemistry research background (consisting of a research question and/or a back-
ground survey), without limitation on the domain of the research question? After
extensive discussions with chemistry experts, we propose an assumption that a
majority of chemistry hypotheses can result from a research background and sev-
eral inspirations. With this key insight, we break the central question into three
smaller fundamental questions. In brief, they are: (1) given a background ques-
tion, whether LLMs can retrieve good inspirations; (2) with background and inspi-
rations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify
good hypotheses to rank them higher. To investigate these questions, we construct
a benchmark consisting of 51 chemistry papers published in Nature, Science, or
a similar level in 2024 (all papers are only available online since 2024). Every
paper is divided by chemistry PhD students into three components: background,
inspirations, and hypothesis. The goal is to rediscover the hypothesis, given only
the background and a large randomly selected chemistry literature corpus con-
sisting of the ground truth inspiration papers, with LLMs trained with data up to
2023. We also develop an LLM-based multi-agent framework 1 that leverages the
assumption, consisting of three stages reflecting the three smaller questions. The
proposed method can rediscover many hypotheses with very high similarity with
the ground truth ones, covering the main innovations.
title: Large (Vision) Language Models are Unsupervised In-Context Learners
abstract: Large (Vision) Language Models are Unsupervised In-Context Learners
Recent advances in large language and vision-language models have enabled zero-
shot inference, allowing models to solve new tasks without task-specific training.
Various adaptation techniques such as prompt engineering, In-Context Learning
(ICL), and supervised fine-tuning can further enhance the model’s performance
on a downstream task, but they require substantial manual effort to construct ef-
In this work, we introduce a joint infer-
fective prompts or labeled examples.
ence framework for fully unsupervised adaptation, eliminating the need for man-
ual prompt engineering and labeled examples. Unlike zero-shot inference, which
makes independent predictions, the joint inference makes predictions simultane-
ously for all inputs in a given task. Since direct joint inference involves compu-
tationally expensive optimization, we develop efficient approximation techniques,
leading to two unsupervised adaptation methods: unsupervised fine-tuning and
unsupervised ICL. We demonstrate the effectiveness of our methods across di-
verse tasks and models, including language-only Llama-3.1 on natural language
processing tasks, reasoning-oriented Qwen2.5-Math on grade school math prob-
lems, vision-language OpenFlamingo on vision tasks, and the API-only access
GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate
substantial improvements over the standard zero-shot approach, including 39%
absolute improvement on the challenging GSM8K math reasoning dataset. Re-
markably, despite being fully unsupervised, our framework often performs on par
with supervised approaches that rely on ground truth labels.
title: Chunk-Distilled Language Modeling
abstract: Chunk-Distilled Language Modeling
We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to
text generation that addresses two challenges in current large language models
(LLMs): the inefficiency of token-level generation, and the difficulty of adapting
to new data and knowledge. Our method combines deep network-based LLMs
with a straightforward retrieval module, which allows the generation of multi-
token text chunks at a single decoding step. Our retrieval framework enables
flexible construction of model- or domain-specific datastores, either leveraging
the internal knowledge of existing models, or incorporating expert insights from
human-annotated corpora. This adaptability allows for enhanced control over
the language model’s distribution without necessitating additional training. We
present the CD-LM formulation along with performance metrics demonstrating
its ability to improve language model performance and efficiency across a diverse
set of downstream applications.1
title: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
abstract: Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
When solving challenging problems, language models (LMs) are able to identify
relevant information from long and complicated contexts. To study how LMs solve
retrieval tasks in diverse situations, we introduce ORION, a collection of structured
retrieval tasks, from text understanding to coding. We apply causal analysis on
ORION for 18 open-source language models with sizes ranging from 125 million
to 70 billion parameters. We find that LMs internally decompose retrieval tasks in
a modular way: middle layers at the last token position process the request, while
late layers retrieve the correct entity from the context. Building on our high-level
understanding, we demonstrate a proof of concept application for scalable internal
oversight of LMs to mitigate prompt-injection while requiring human supervision
on only a single input.
title: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
abstract: DiscoveryBench: Towards Data-Driven Discovery with Large Language Models
Can the rapid advances in code generation, function calling, and data analysis
using large language models (LLMs) help automate the search and verification of
hypotheses purely from a set of provided datasets? To evaluate this question, we
present DISCOVERYBENCH, the first comprehensive benchmark that formalizes
the multi-step process of data-driven discovery. The benchmark is designed to
systematically assess current model capabilities in discovery tasks and provide a
useful resource for improving them. Our benchmark contains 264 tasks collected
across 6 diverse domains, such as sociology and engineering, by manually deriving
discovery workflows from published papers to approximate the real-world chal-
lenges faced by researchers, where each task is defined by a dataset, its metadata,
and a discovery goal in natural language. We additionally provide 903 synthetic
tasks to conduct controlled evaluations on data-driven workflows that are not cov-
ered in the manually collected split. Furthermore, our structured formalism of
data-driven discovery enables a facet-based evaluation that provides useful insights
into different failure modes. We evaluate several popular LLM-based reasoning
frameworks using both open and closed LLMs as baselines on DISCOVERYBENCH
and find that even the best system scores only 25%. Our benchmark, thus, illus-
trates the challenges in autonomous data-driven discovery and serves as a valuable
resource for the community to make progress.
title: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
abstract: Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents
Effective research ideation is a critical step for scientific research. However, the
exponential increase in scientific literature makes it challenging for researchers to
stay current with recent advances and identify meaningful research directions. Re-
cent developments in large language models (LLMs) suggest a promising avenue
for automating the generation of novel research ideas. However, existing methods
for idea generation either trivially prompt LLMs or directly expose LLMs to ex-
tensive literature without indicating useful information. Inspired by the research
process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-
based agent that organizes relevant literature in a chain structure to effectively
mirror the progressive development in a research domain. This organization facil-
itates LLMs to capture the current advancements in research, thereby enhancing
their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation
protocol that can comprehensively evaluate idea generation methods from dif-
ferent perspectives, aligning closely with the preferences of human researchers.
Experimental results indicate that the CoI agent consistently outperforms other
methods and shows comparable quality as humans in research idea generation.
Moreover, our CoI agent is budget-friendly, with a minimum cost of $0.50 to gen-
erate a candidate idea and its corresponding experimental design1.
title: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
abstract: Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Large language models (LLMs) have significantly benefited from training on di-
verse, high-quality task-specific data, leading to impressive performance across
a range of downstream applications. Current methods often rely on human-
annotated data or predefined task templates to direct powerful LLMs in synthe-
sizing task-relevant data for effective model training. However, this dependence
on manually designed components may constrain the scope of generated data,
potentially overlooking critical edge cases or novel scenarios that could chal-
lenge the model.
In this paper, we present a novel approach, REVERSEGEN,
designed to automatically generate effective training samples that expose the
weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained
to produce queries that lead target models to generate unsatisfactory responses.
These failure-inducing queries are then used to construct training data, helping
to address the models’ shortcomings and improve overall performance. Our ap-
proach is flexible and can be applied to models of various scales (3B, 7B, and
8B). We evaluate REVERSEGEN on three key applications—safety, honesty, and
math—demonstrating that our generated data is both highly effective and diverse.
Models fine-tuned with REVERSEGEN-generated data consistently outperform
those trained on human-annotated or general model-generated data, offering a new
perspective on data synthesis for task-specific LLM enhancement. 1.
title: Planning in Natural Language Improves LLM Search for Code Generation
abstract: Planning in Natural Language Improves LLM Search for Code Generation
While scaling training compute has led to remarkable improvements in large
language models (LLMs), scaling inference compute only recently began to yield
analogous gains. We hypothesize that a core missing component is a lack of diverse
LLM outputs, leading to inefficient search due to models repeatedly sampling
highly similar, yet incorrect generations. We empirically demonstrate that this
lack of diversity can be mitigated by searching over candidate plans for solving a
problem in natural language. Based on this insight, we propose PLANSEARCH, a
novel search algorithm which shows strong results across HumanEval+, MBPP+,
and LiveCodeBench (a contamination-free benchmark for competitive coding).
PLANSEARCH generates a diverse set of observations about the problem and uses
these observations to construct plans for solving the problem. By searching over
plans in natural language rather than directly over code solutions, PLANSEARCH
explores a significantly more diverse range of potential solutions compared to
baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet
achieves a pass@200 of 77.0% on LiveCodeBench, outperforming both the best
pass-rate achieved without any search (pass@1 = 41.4%) and using standard
repeated sampling on top of existing non-search models (pass@200 = 60.6%).
Finally, we show that, across all models, search algorithms, and benchmarks
analyzed, we can accurately predict performance gains from search as a function
of the diversity over generated ideas.
You should generate 1 project proposal(s) on this topic. Be creative and diverse in the idea generation. The above papers are only for inspiration and you should not just make some incremental modifications on top of them. Instead, you should make sure your ideas are novel and distinct from the prior literature. Each project proposal should be described as: (1) Problem Statement: State the problem statement, which should be closely related to the topic description and something that is not well solved yet. (2) Motivation: Explain the inspiration of the proposed method and why it would work well. (3) Proposed Method: Propose your new method and describe it in detail. The proposed method should be maximally different from all existing work and baselines, and be more advanced and effective than the baselines. You should be as creative as possible in proposing new methods. Make sure to write down the idea as a list of atomic steps where each step is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it. (4) Experiment Plan: Specify all the experiment steps, baselines, and evaluation metrics. If using existing datasets, mention the names of the datasets; or alternatively, explain how to construct the datasets. Make sure to write this section as list of different experiments to do where each one is described in detail with concrete action items and is at least 2 sentences long. Avoid vague steps and use very specific terms to tell the student what to implement and how to do it.
Focus on proposing novel empirical methods. You are encouraged to use a diverse set of techniques or their combinations. The proposed method section should specify all the details involved, such as how to get the data, what's the training objective, how to construct the prompts, all the datasets and metrics, etc. You should aim for projects that can potentially win best paper awards at top AI and LLM conferences like NeurIPS and ICLR.
Output the proposals in json format as a dictionary, where you should generate a short idea name as the key and the actual idea description as the value. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.