ID
large_stringlengths 10
61
| year
int64 1.96k
2.03k
| title
large_stringlengths 4
560
| abstract
large_stringlengths 0
12.8k
|
|---|---|---|---|
liu-etal-2025-learning | 2,025 | Learning to Refuse: Towards Mitigating Privacy Risks in LLMs | Large language models (LLMs) exhibit remarkable capabilities in understanding and generating natural language. However, these models can inadvertently memorize private information, posing significant privacy risks. This study addresses the challenge of enabling LLMs to protect specific individuals' private data without the need for complete retraining. We propose RETURN, a Real-world pErsonal daTa UnleaRNing dataset, comprising 2,492 individuals from Wikipedia with associated QA pairs, to evaluate machine unlearning (MU) methods for protecting personal data in a realistic scenario. Additionally, we introduce the Name-Aware Unlearning Framework (NAUF) for Privacy Protection, which enables the model to learn which individuals' information should be protected without affecting its ability to answer questions related to other unrelated individuals. Our extensive experiments demonstrate that NAUF achieves a state-of-the-art average unlearning score, surpassing the best baseline method by 5.65 points, effectively protecting target individuals' personal data while maintaining the model`s general capabilities. |
qiang-etal-2025-exploring | 2,025 | Exploring Unified Training Framework for Multimodal User Profiling | With the emergence of social media and e-commerce platforms, accurate user profiling has become increasingly vital for recommendation systems and personalized services. Recent studies have focused on generating detailed user profiles by extracting various aspects of user attributes from textual reviews. Nevertheless, these investigations have not fully exploited the potential of the abundant multimodal data at hand. In this study, we propose a novel task called multimodal user profiling. This task emphasizes the utilization of both review texts and their accompanying images to create comprehensive user profiles. By integrating textual and visual data, we leverage their complementary strengths, enabling the generation of more holistic user representations. Additionally, we explore a unified joint training framework with various multimodal training strategies that incorporate users' historical review texts and images for user profile generation. Our experimental results underscore the significance of multimodal data in enhancing user profile generation and demonstrate the effectiveness of the proposed unified joint training approach. |
goto-etal-2025-acquiring | 2,025 | Acquiring Bidirectionality via Large and Small Language Models | Using token representation from bidirectional language models (LMs) such as BERT is still a widely used approach for token-classification tasks. Even though there exist much larger unidirectional LMs such as Llama-2, they are rarely used to replace the token representation of bidirectional LMs. In this work, we hypothesize that their lack of bidirectionality is what is keeping unidirectional LMs behind. To that end, we propose to newly train a small backward LM and concatenate its representations to those of an existing LM for downstream tasks. Through experiments in token-classification tasks, we demonstrate that introducing backward model can improve the benchmark performance by more than 10 points. Furthermore, we show that the proposed method is especially effective for rare domains and in few-shot learning settings. |
li-etal-2025-enhancing | 2,025 | Enhancing One-Shot Pruned Pre-trained Language Models through Sparse-Dense-Sparse Mechanism | Pre-trained language models (PLMs) are engineered to be robust in contextual understanding and exhibit outstanding performance in various natural language processing tasks. However, their considerable size incurs significant computational and storage costs. Modern pruning strategies employ retraining-free one-shot techniques to compress PLMs; however, these approaches often lead to an indispensable reduction in performance. In this paper, we propose SDS, a Sparse-Dense-Sparse pruning framework to enhance the performance of the pruned PLMs from a weight distribution optimization perspective. We outline the pruning process in three steps. Initially, we prune less critical connections in the model using conventional one-shot pruning methods. Next, we reconstruct a dense model featuring a pruning-friendly weight distribution by reactivating pruned connections with sparse regularization. Finally, we perform a second pruning round, yielding a superior pruned model compared to the initial pruning. Experiments demonstrate that SDS outperforms the state-of-the-art pruning techniques SparseGPT and Wanda under an identical sparsity configuration. For instance, SDS reduces perplexity by 5.16 on Raw-Wikitext2 and improves average accuracy by 3.86{\%} across multiple zero-shot benchmarks for LLaMA-3-8B compared to Wanda with 2:4 sparsity. |
he-etal-2025-language | 2,025 | Language Models over Large-Scale Knowledge Base: on Capacity, Flexibility and Reasoning for New Facts | Advancements in language models (LMs) have sparked interest in exploring their potential as knowledge bases (KBs) due to their high capability for storing huge amounts of factual knowledge and semantic understanding. However, existing studies face challenges in quantifying the extent of large-scale knowledge packed into LMs and lack systematic studies on LMs' structured reasoning capabilities over the infused knowledge. Addressing these gaps, our research investigates whether LMs can effectively act as large-scale KBs after training over an expansive set of world knowledge triplets via addressing the following three crucial questions: (1) How do LMs of different sizes perform at storing world knowledge of different frequencies in a large-scale KB? (2) How flexible are these LMs in recalling the stored knowledge when prompted with natural language queries? (3) After training on the abundant world knowledge, can LMs additionally gain the ability to reason over such information to infer new facts? Our findings indicate that while medium-scaled LMs hold promise as world knowledge bases capable of storing and responding with flexibility, enhancements in their reasoning capabilities are necessary to fully realize their potential. |
guo-etal-2025-multi | 2,025 | Multi-View Incongruity Learning for Multimodal Sarcasm Detection | Multimodal sarcasm detection (MSD) is essential for various downstream tasks. Existing MSD methods tend to rely on spurious correlations. These methods often mistakenly prioritize non-essential features yet still make correct predictions, demonstrating poor generalizability beyond training environments. Regarding this phenomenon, this paper undertakes several initiatives. Firstly, we identify two primary causes that lead to the reliance of spurious correlations. Secondly, we address these challenges by proposing a novel method that integrate Multimodal Incongruities via Contrastive Learning (MICL) for multimodal sarcasm detection. Specifically, we first leverage incongruity to drive multi-view learning from three views: token-patch, entity-object, and sentiment. Then, we introduce extensive data augmentation to mitigate the biased learning of the textual modality. Additionally, we construct a test set, SPMSD, which consists potential spurious correlations to evaluate the the model`s generalizability. Experimental results demonstrate the superiority of MICL on benchmark datasets, along with the analyses showcasing MICL`s advancement in mitigating the effect of spurious correlation. |
mina-etal-2025-cognitive | 2,025 | Cognitive Biases, Task Complexity, and Result Intepretability in Large Language Models | In humans, cognitive biases are systematic deviations from rationality in judgment that simplify complex decisions. They typically manifest as a consequence of learned behaviors or limitations on information processing capabilities. Recent work has shown that these biases can percolate through training data and ultimately be learned by language models. We examine different groups of models, factoring in model size and type (base or instructed) for four kinds of cognitive bias: primacy, recency, common token, and majority class bias. We evaluate the performance of each model for each type of bias in different settings using simple and complex variants of datasets. Our results show that some biases have much stronger effects than others, and that task complexity plays a part in eliciting stronger effects for some of these biases as measured by effect size. We show that some cognitive biases such as common token and majority class bias are not straightforward to evaluate, and that, contrary to some of the previous literature, some effects that have been previously classified as common token bias in the literature are actually due to primacy and recency bias. |
satheesh-etal-2025-robustness | 2,025 | Robustness Evaluation of the German Extractive Question Answering Task | To ensure reliable performance of Question Answering (QA) systems, evaluation of robustness is crucial. Common evaluation benchmarks commonly only include performance metrics, such as Exact Match (EM) and the F1 score. However, these benchmarks overlook critical factors for the deployment of QA systems. This oversight can result in systems vulnerable to minor perturbations in the input such as typographical errors. While several methods have been proposed to test the robustness of QA models, there has been minimal exploration of these approaches for languages other than English. This study focuses on the robustness evaluation of German language QA models, extending methodologies previously applied primarily to English. The objective is to nurture the development of robust models by defining an evaluation method specifically tailored to the German language. We assess the applicability of perturbations used in English QA models for German and perform a comprehensive experimental evaluation with eight models. The results show that all models are vulnerable to character-level perturbations. Additionally, the comparison of monolingual and multilingual models suggest that the former are less affected by character and word-level perturbations. |
xu-etal-2025-enhancing | 2,025 | Enhancing Multimodal Named Entity Recognition through Adaptive Mixup Image Augmentation | Multimodal named entity recognition (MNER) extends traditional named entity recognition (NER) by integrating visual and textual information. However, current methods still face significant challenges due to the text-image mismatch problem. Recent advancements in text-to-image synthesis provide promising solutions, as synthesized images can introduce additional visual context to enhance MNER model performance. To fully leverage the benefits of both original and synthesized images, we propose an adaptive mixup image augmentation method. This method generates augmented images by determining the mixing ratio based on the matching score between the text and image, utilizing a triplet loss-based Gaussian Mixture Model (TL-GMM). Our approach is highly adaptable and can be seamlessly integrated into existing MNER models. Extensive experiments demonstrate consistent performance improvements, and detailed ablation studies and case studies confirm the effectiveness of our method. |
zhao-etal-2025-bridging | 2,025 | Bridging Modality Gap for Effective Multimodal Sentiment Analysis in Fashion-related Social Media | Multimodal sentiment analysis for fashion-related social media is essential for understanding how consumers appraise fashion products across platforms like Instagram and Twitter, where both textual and visual elements contribute to sentiment expression. However, a notable challenge in this task is the modality gap, where the different information density between text and images hinders effective sentiment analysis. In this paper, we propose a novel multimodal framework that addresses this challenge by introducing pseudo data generated by a two-stage framework. We further utilize a multimodal fusion approach that efficiently integrates the information from various modalities for sentiment classification of fashion posts. Experiments conducted on a comprehensive dataset demonstrate that our framework significantly outperforms existing unimodal and multimodal baselines, highlighting its effectiveness in bridging the modality gap for more accurate sentiment classification in fashion-related social media posts. |
van-noord-etal-2025-quality | 2,025 | Quality Beyond A Glance: Revealing Large Quality Differences Between Web-Crawled Parallel Corpora | Parallel corpora play a vital role in advanced multilingual natural language processing tasks, notably in machine translation (MT). The recent emergence of numerous large parallel corpora, often extracted from multilingual documents on the Internet, has expanded the available resources. Nevertheless, the quality of these corpora remains largely unexplored, while there are large differences in how the corpora are constructed. Moreover, how the potential differences affect the performance of neural MT (NMT) systems has also received limited attention. This study addresses this gap by manually and automatically evaluating four well-known publicly available parallel corpora across eleven language pairs. Our findings are quite concerning: all corpora contain a substantial amount of noisy sentence pairs, with CCMatrix and CCAligned having well below of 50{\%} reasonably clean pairs. MaCoCu and ParaCrawl generally have higher quality texts, though around a third of the texts still have clear issues. While corpus size impacts NMT models' performance, our study highlights the critical role of quality: higher-quality corpora consistently yield better-performing NMT models when controlling for size. |
bao-etal-2025-mllm | 2,025 | MLLM-I2W: Harnessing Multimodal Large Language Model for Zero-Shot Composed Image Retrieval | Combined Image Retrieval (CIR) involves retrieving an image based on a reference image and a brief text description, which is widely present in various scenarios such as fashion recommendation. Existing methods can be mainly divided into two categories, respectively supervised CIR methods and Zero-Shot CIR (ZS-CIR) methods. In contrast to supervised CIR methods, which need manually annotated triples for training task-specific models, ZS-CIR models can be trained using images datasets only and performs well. However, ZS-CIR still faces the primary challenge of learning how to map pseudo-words to images within the joint image-text embedding space. Therefore, in this paper, we propose a novel image-text mapping network, named MLLM-I2W, which adaptively converts description-related image information into pseudo-word markers for precise ZS-CIR. Specifically, the image and text encoding enhancement module within the MLLM prompt selects subject headings and generates text descriptions. It then reduces the modality gap between images and text using uncertainty modeling. An adaptive weighting module and a prototype are proposed to adjust and learn the deep fusion features, which are further mapped to pseudo-word markers via well-designed MOE-based mapping network. Our model demonstrates consistent improvements across common CIR benchmarks, including COCO, CIRR, and Fashion-IQ. |
heitz-etal-2025-linguistic | 2,025 | Linguistic Features Extracted by GPT-4 Improve Alzheimer`s Disease Detection based on Spontaneous Speech | Alzheimer`s Disease (AD) is a significant and growing public health concern. Investigating alterations in speech and language patterns offers a promising path towards cost-effective and non-invasive early detection of AD on a large scale. Large language models (LLMs), such as GPT, have enabled powerful new possibilities for semantic text analysis. In this study, we leverage GPT-4 to extract five semantic features from transcripts of spontaneous patient speech. The features capture known symptoms of AD, but they are difficult to quantify effectively using traditional methods of computational linguistics. We demonstrate the clinical significance of these features and further validate one of them ({\textquotedblleft}Word-Finding Difficulties{\textquotedblright}) against a proxy measure and human raters. When combined with established linguistic features and a Random Forest classifier, the GPT-derived features significantly improve the detection of AD. Our approach proves effective for both manually transcribed and automatically generated transcripts, representing a novel and impactful use of recent advancements in LLMs for AD speech analysis. |
kuribayashi-baldwin-2025-vision | 2,025 | Does Vision Accelerate Hierarchical Generalization in Neural Language Learners? | Neural language models (LMs) are arguably less data-efficient than humans from a language acquisition perspective. One fundamental question is why this human{--}LM gap arises. This study explores the advantage of \textit{grounded} language acquisition, specifically the impact of visual information {---} which humans can usually rely on but LMs largely do not have access to during language acquisition {---} on syntactic generalization in LMs. Our experiments, following the poverty of stimulus paradigm under two scenarios (using artificial vs. naturalistic images), demonstrate that if the alignments between the linguistic and visual components are clear in the input, access to vision data does help with the syntactic generalization of LMs, but if not, visual input does not help. This highlights the need for additional biases or signals, such as mutual gaze, to enhance cross-modal alignment and enable efficient syntactic generalization in multimodal LMs. |
hosseini-etal-2025-efficient | 2,025 | Efficient Solutions For An Intriguing Failure of LLMs: Long Context Window Does Not Mean LLMs Can Analyze Long Sequences Flawlessly | Large Language Models (LLMs) have demonstrated remarkable capabilities in comprehending and analyzing lengthy sequential inputs, owing to their extensive context windows that allow processing millions of tokens in a single forward pass. However, this paper uncovers a surprising limitation: LLMs fall short when handling long input sequences. We investigate this issue using three datasets and two tasks (sentiment analysis and news categorization) across various LLMs, including Claude 3, Gemini Pro, GPT 3.5 Turbo, Llama 3 Instruct, and Mistral Instruct models. To address this limitation, we propose and evaluate ad-hoc solutions that substantially enhance LLMs' performance on long input sequences by up to 50{\%}, while reducing API cost and latency by up to 93{\%} and 50{\%}, respectively. |
zhang-long-2025-mld | 2,025 | MLD-EA: Check and Complete Narrative Coherence by Introducing Emotions and Actions | Narrative understanding and story generation are critical challenges in natural language processing (NLP), with much of the existing research focused on summarization and question-answering tasks. While previous studies have explored predicting plot endings and generating extended narratives, they often neglect the logical coherence within stories, leaving a significant gap in the field. To address this, we introduce the Missing Logic Detector by Emotion and Action (MLD-EA) model, which leverages large language models (LLMs) to identify narrative gaps and generate coherent sentences that integrate seamlessly with the story`s emotional and logical flow. The experimental results demonstrate that the MLD-EA model enhances narrative understanding and story generation, highlighting LLMs' potential as effective logic checkers in story writing with logical coherence and emotional consistency. This work fills a gap in NLP research and advances border goals of creating more sophisticated and reliable story-generation systems. |
tsuji-etal-2025-subregweigh | 2,025 | SubRegWeigh: Effective and Efficient Annotation Weighing with Subword Regularization | NLP datasets may still contain annotation errors, even when they are manually annotated. Researchers have attempted to develop methods to automatically reduce the adverse effect of errors in datasets. However, existing methods are time-consuming because they require many trained models to detect errors. This paper proposes a time-saving method that utilizes a tokenization technique called subword regularization to simulate multiple error detection models for detecting errors. Our proposed method, SubRegWeigh, can perform annotation weighting four to five times faster than the existing method. Additionally, SubRegWeigh improved performance in document classification and named entity recognition tasks. In experiments with pseudo-incorrect labels, SubRegWeigh clearly identifies pseudo-incorrect labels as annotation errors. Our code is available at https://github.com/4ldk/SubRegWeigh. |
yang-etal-2025-rethinking | 2,025 | Rethinking Long Context Generation from the Continual Learning Perspective | Due to the limited context window, Large Language Models (LLMs) struggle with processing long contexts. Although fine-tuning can extend the context window, it incurs substantial computation costs. In contrast, recent tuning-free approaches reallocate the attention mechanism or incorporate temporary trainable parameters. In this work, by jointly modeling instance-level generation with a limited context window and learning over sequential data, we rethink the long context generation of LLMs from a continual learning perspective. In practice, we inspect existing representative approaches and analyze their synergy with continual learning strategies. Moreover, we integrate these strategies into current approaches to further boost LLMs' efficiency in processing long contexts. Comprehensive experiments and analysis confirm the feasibility of continual learning insights for improving long-context processing. |
wang-etal-2025-ltrs | 2,025 | LTRS: Improving Word Sense Disambiguation via Learning to Rank Senses | Word Sense Disambiguation (WSD) is a fundamental task critical for accurate semantic understanding. Conventional training strategies usually only consider predefined senses for target words and learn each of them from relatively limited instances, neglecting the influence of similar ones. To address these problems, we propose the method of Learning to Rank Senses (LTRS) to enhance the task. This method helps a model learn to represent and disambiguate senses from a broadened range of instances via ranking an expanded list of sense definitions. By employing LTRS, our model achieves a SOTA F1 score of 79.6{\%} in Chinese WSD and exhibits robustness in low-resource settings. Moreover, it shows excellent training efficiency, achieving faster convergence than previous methods. This provides a new technical approach to WSD and may also apply to the task for other languages. |
galletti-etal-2025-keywords | 2,025 | Are Your Keywords Like My Queries? A Corpus-Wide Evaluation of Keyword Extractors with Real Searches | Keyword Extraction (KE) is essential in Natural Language Processing (NLP) for identifying key terms that represent the main themes of a text, and it is vital for applications such as information retrieval, text summarisation, and document classification. Despite the development of various KE methods {---} including statistical approaches and advanced deep learning models {---} evaluating their effectiveness remains challenging. Current evaluation metrics focus on keyword quality, balance, and overlap with annotations from authors and professional indexers, but neglect real-world information retrieval needs. This paper introduces a novel evaluation method designed to overcome this limitation by using real query data from Google Trends and can be used with both supervised and unsupervised KE approaches. We applied this method to three popular KE approaches (YAKE, RAKE and KeyBERT) and found that KeyBERT was the most effective in capturing users' top queries, with RAKE also showing surprisingly good performance. The code is open-access and publicly available. |
loredo-lopez-etal-2025-nyt | 2,025 | NYT-Connections: A Deceptively Simple Text Classification Task that Stumps System-1 Thinkers | Large Language Models (LLMs) have shown impressive performance on various benchmarks, yet their ability to engage in deliberate reasoning remains questionable. We present NYT-Connections, a collection of 358 simple word classification puzzles derived from the New York Times Connections game. This benchmark is designed to penalize quick, intuitive {\textquotedblleft}System 1{\textquotedblright} thinking, isolating fundamental reasoning skills. We evaluated six recent LLMs, a simple machine learning heuristic, and humans across three configurations: single-attempt, multiple attempts without hints, and multiple attempts with contextual hints. Our findings reveal a significant performance gap: even top-performing LLMs like GPT-4 fall short of human performance by nearly 30{\%}. Notably, advanced prompting techniques such as Chain-of-Thought and Self-Consistency show diminishing returns as task difficulty increases. NYT-Connections uniquely combines linguistic isolation, resistance to intuitive shortcuts, and regular updates to mitigate data leakage, offering a novel tool for assessing LLM reasoning capabilities. |
basar-etal-2025-well | 2,025 | How Well Can Large Language Models Reflect? A Human Evaluation of LLM-generated Reflections for Motivational Interviewing Dialogues | Motivational Interviewing (MI) is a counseling technique that promotes behavioral change through reflective responses to mirror or refine client statements. While advanced Large Language Models (LLMs) can generate engaging dialogues, challenges remain for applying them in a sensitive context such as MI. This work assesses the potential of LLMs to generate MI reflections via three LLMs: GPT-4, Llama-2, and BLOOM, and explores the effect of dialogue context size and integration of MI strategies for reflection generation by LLMs. We conduct evaluations using both automatic metrics and human judges on four criteria: appropriateness, relevance, engagement, and naturalness, to assess whether these LLMs can accurately generate the nuanced therapeutic communication required in MI. While we demonstrate LLMs' potential in generating MI reflections comparable to human therapists, content analysis shows that significant challenges remain. By identifying the strengths and limitations of LLMs in generating empathetic and contextually appropriate reflections in MI, this work contributes to the ongoing dialogue in enhancing LLM`s role in therapeutic counseling. |
sun-etal-2025-rethinking | 2,025 | Rethinking the Alignment of Psychotherapy Dialogue Generation with Motivational Interviewing Strategies | Recent advancements in large language models (LLMs) have shown promise in generating psychotherapeutic dialogues, particularly in the context of motivational interviewing (MI). However, the inherent lack of transparency in LLM outputs presents significant challenges given the sensitive nature of psychotherapy. Applying MI strategies, a set of MI skills, to generate more controllable therapeutic-adherent conversations with explainability provides a possible solution. In this work, we explore the alignment of LLMs with MI strategies by first prompting the LLMs to predict the appropriate strategies as reasoning and then utilizing these strategies to guide the subsequent dialogue generation. We seek to investigate whether such alignment leads to more controllable and explainable generations. Multiple experiments including automatic and human evaluations are conducted to validate the effectiveness of MI strategies in aligning psychotherapy dialogue generation. Our findings demonstrate the potential of LLMs in producing strategically aligned dialogues and suggest directions for practical applications in psychotherapeutic settings. |
kumar-etal-2025-enhancing | 2,025 | Enhancing Zero-shot Chain of Thought Prompting via Uncertainty-Guided Strategy Selection | Chain-of-thought (CoT) prompting has significantly enhanced the the capability of large language models (LLMs) by structuring their reasoning processes. However, existing methods face critical limitations: handcrafted demonstrations require extensive human expertise, while trigger phrases are prone to inaccuracies. In this paper, we propose the Zero-shot Uncertainty-based Selection (ZEUS) method, a novel approach that improves CoT prompting by utilizing uncertainty estimates to select effective demonstrations without needing access to model parameters. Unlike traditional methods, ZEUS offers high sensitivity in distinguishing between helpful and ineffective questions, ensuring more precise and reliable selection. Our extensive evaluation shows that ZEUS consistently outperforms existing CoT strategies across four challenging reasoning benchmarks, demonstrating its robustness and scalability. |
feng-etal-2025-word | 2,025 | Word-level Cross-lingual Structure in Large Language Models | Large Language Models (LLMs) have demonstrated exceptional performance across a broad spectrum of cross-lingual Natural Language Processing (NLP) tasks. However, previous methods predominantly focus on leveraging parallel corpus to conduct instruction data for continuing pre-training or fine-tuning. They ignored the state of parallel data on the hidden layers of LLMs. In this paper, we demonstrate Word-level Cross-lingual Structure (WCS) of LLM which proves that the word-level embedding on the hidden layers are isomorphic between languages. We find that the hidden states of different languages' input on the LLMs hidden layers can be aligned with an orthogonal matrix on word-level. We prove this conclusion in both mathematical and downstream task ways on two representative LLM foundations, LLaMA2 and BLOOM. Besides, we propose an Isomorphism-based Data Augmentation (IDA) method to apply the WCS on a downstream cross-lingual task, Bilingual Lexicon Induction (BLI), in both supervised and unsupervised ways. The experiment shows the significant improvement of our proposed method over all the baselines, especially on low-resource languages. |
zhang-etal-2025-trucidator | 2,025 | Trucidator: Document-level Event Factuality Identification via Hallucination Enhancement and Cross-Document Inference | Document-level event factuality identification (DEFI) assesses the veracity degree to which an event mentioned in a document has happened, which is crucial for many natural language processing tasks. Previous work assesses event factuality by solely relying on the semantic information within a single document, which fails to identify hard cases where the document itself is hallucinative or counterfactual. There is also a pressing need for more suitable data of this kind. To tackle these issues, we construct Factualusion, a novel corpus with hallucination features that can be used not only for DEFI but can also be applied for hallucination evaluation for large language models. We further propose Trucidator, a graph-based framework that constructs intra-document and cross-document graphs and employs a multi-task learning paradigm to acquire more robust node embeddings, leveraging cross-document inference for more accurate identification. Experiments show that our proposed framework outperformed several baselines, demonstrating the effectiveness of our method. |
avram-etal-2025-rolargesum | 2,025 | RoLargeSum: A Large Dialect-Aware Romanian News Dataset for Summary, Headline, and Keyword Generation | Using supervised automatic summarisation methods requires sufficient corpora that include pairs of documents and their summaries. Similarly to many tasks in natural language processing, most of the datasets available for summarization are in English, posing challenges for developing summarization models in other languages. Thus, in this work, we introduce RoLargeSum, a novel large-scale summarization dataset for the Romanian language crawled from various publicly available news websites from Romania and the Republic of Moldova that were thoroughly cleaned to ensure a high-quality standard. RoLargeSum contains more than 615K news articles, together with their summaries, as well as their headlines, keywords, dialect, and other metadata that we found on the targeted websites. We further evaluated the performance of several BART variants and open-source large language models on RoLargeSum for benchmarking purposes. We manually evaluated the results of the best-performing system to gain insight into the potential pitfalls of this data set and future development. |
di-bonaventura-etal-2025-detection | 2,025 | From Detection to Explanation: Effective Learning Strategies for LLMs in Online Abusive Language Research | Abusive language detection relies on understanding different levels of intensity, expressiveness and targeted groups, which requires commonsense reasoning, world knowledge and linguistic nuances that evolve over time. Here, we frame the problem as a knowledge-guided learning task, and demonstrate that LLMs' implicit knowledge without an accurate strategy is not suitable for multi-class detection nor explanation generation. We publicly release GLlama Alarm, the knowledge-Guided version of Llama-2 instruction fine-tuned for multi-class abusive language detection and explanation generation. By being fine-tuned on structured explanations and external reliable knowledge sources, our model mitigates bias and generates explanations that are relevant to the text and coherent with human reasoning, with an average 48.76{\%} better alignment with human judgment according to our expert survey. |
ravikiran-etal-2025-teemil | 2,025 | TEEMIL : Towards Educational MCQ Difficulty Estimation in Indic Languages | Difficulty estimation of multiple-choice questions (MCQs) is crucial for creating effective educational assessments, yet remains underexplored in Indic languages like Hindi and Kannada due to the lack of comprehensive datasets. This paper addresses this gap by introducing two datasets, TEEMIL-H and TEEMIL-K, containing 4689 and 4215 MCQs, respectively, with manually annotated difficulty labels. We benchmark these datasets using state-of-the-art multilingual models and conduct ablation studies to analyze the effect of context, the impact of options, and the presence of the None of the Above (NOTA) option on difficulty estimation. Our findings establish baselines for difficulty estimation in Hindi and Kannada, offering valuable insights into improving model performance and guiding future research in MCQ difficulty estimation . |
kirstein-etal-2025-whats | 2,025 | What`s Wrong? Refining Meeting Summaries with LLM Feedback | Meeting summarization has become a critical task since digital encounters have become a common practice. Large language models (LLMs) show great potential in summarization, offering enhanced coherence and context understanding compared to traditional methods. However, they still struggle to maintain relevance and avoid hallucination. We introduce a multi-LLM correction approach for meeting summarization using a two-phase process that mimics the human review process: mistake identification and summary refinement. We release QMSum Mistake, a dataset of 200 automatically generated meeting summaries annotated by humans on nine error types, including structural, omission, and irrelevance errors. Our experiments show that these errors can be identified with high accuracy by an LLM. We transform identified mistakes into actionable feedback to improve the quality of a given summary measured by relevance, informativeness, conciseness, and coherence. This post-hoc refinement effectively improves summary quality by leveraging multiple LLMs to validate output quality. Our multi-LLM approach for meeting summarization shows potential for similar complex text generation tasks requiring robustness, action planning, and discussion towards a goal. |
sun-etal-2025-scene | 2,025 | Scene Graph and Dependency Grammar Enhanced Remote Sensing Change Caption Network (SGD-RSCCN) | With the continuous advancement of remote sensing technology, it is easier to obtain high-resolution, multi-temporal and multi-spectral images. The images carry rich information of ground objects. However, how to effectively extract useful information from the complex image data and convert it into understandable semantic descriptions remains a challenge. To deal with the challenges, we propose a Scene Graph and Dependency Grammar Enhanced Remote Sensing Change Caption Network (SGD-RSCCN) to improve the accuracy and naturalness of extracting and describing change information from remote sensing images. By combining advanced visual analysis technology and natural language processing technology, the network not only optimizes the problem of insufficient understanding of complex scenes, but also enhances the ability to capture dynamic changes, thereby generating more accurate and smooth natural language description. In addition, we also proposes the decoder based on prior knowledge, which further improves the readability and comprehensibility of the description. Extensive experiments on LEVIR-CC and Dubai-CC datasets verify the advantages of the proposed method in generating accurate and true descriptions. |
ruiz-dolz-etal-2025-looking | 2,025 | Looking at the Unseen: Effective Sampling of Non-Related Propositions for Argument Mining | Traditionally, argument mining research has approached the task of automatic identification of argument structures by using existing definitions of what constitutes an argument, while leaving the equally important matter of what does not qualify as an argument unaddressed. With the ability to distinguish between what is and what is not a natural language argument being at the core of argument mining as a field, it is interesting that no previous work has explored approaches to effectively select non-related propositions (i.e., propositions that are not connected through an argumentative relation, such as support or attack) that improve the data for learning argument mining tasks better. In this paper, we address the question of how to effectively sample non-related propositions from six different argument mining corpora belonging to different domains and encompassing both monologue and dialogue forms of argumentation. To that end, in addition to considering undersampling baselines from previous work, we propose three new sampling strategies relying on context (i.e., short/long) and the semantic similarity between propositions. Our results indicate that using more informed sampling strategies improves the performance, not only when evaluating models on their respective test splits, but also in the case of cross-domain evaluation. |
mei-etal-2025-aligned | 2,025 | \textquotedblleftNot Aligned\textquotedblright is Not \textquotedblleftMalicious\textquotedblright: Being Careful about Hallucinations of Large Language Models' Jailbreak | {\textquotedblleft}Jailbreak{\textquotedblright} is a major safety concern of Large Language Models (LLMs), which occurs when malicious prompts lead LLMs to produce harmful outputs, raising issues about the reliability and safety of LLMs. Therefore, an effective evaluation of jailbreaks is very crucial to develop its mitigation strategies. However, our research reveals that many jailbreaks identified by current evaluations may actually be hallucinations{---}erroneous outputs that are mistaken for genuine safety breaches. This finding suggests that some perceived vulnerabilities might not represent actual threats, indicating a need for more precise red teaming benchmarks. To address this problem, we propose the Benchmark for reliABilitY and jailBreak haLlUcination Evaluation (BabyBLUE). BabyBLUE introduces a specialized validation framework including various evaluators to enhance existing jailbreak benchmarks, ensuring outputs are useful malicious instructions. Additionally, BabyBLUE presents a new dataset as an augmentation to the existing red teaming benchmarks, specifically addressing hallucinations in jailbreaks, aiming to evaluate the true potential of jailbroken LLM outputs to cause harm to human society. |
mikulova-etal-2025-form | 2,025 | From Form to Meaning: The Case of Particles within the Prague Dependency Treebank Annotation Scheme | In the last decades, computational linguistics has become increasingly interested in annotation schemes that aim at an adequate description of the meaning of the sentences and texts. Discussions are ongoing on an appropriate annotation scheme for a large and complex amount of diverse information. In this contribution devoted to description of polyfunctional uninflected words (namely particles), i.e. words which, although having only one paradigmatic form, can have several different syntactic functions and even express relatively different semantic distinctions, we argue that it is the multi-layer system (linked from meaning to text) that allows a comprehensive description of the relations between morphological properties, syntactic function and expressed meaning, and thus contributes to greater accuracy in the description of the phenomena concerned and to the overall consistency of the annotated data. These aspects are demonstrated within the Prague Dependency Treebank annotation scheme, whose pioneering proposal can be found in the first COLING proceedings from 1965 (Sgall 1965), and to this day, the concept has proved to be sound and serves very well for complex annotation. |
lawan-etal-2025-enhancing | 2,025 | Enhancing Long-range Dependency with State Space Model and Kolmogorov-Arnold Networks for Aspect-based Sentiment Analysis | Aspect-based Sentiment Analysis (ABSA) evaluates sentiments toward specific aspects of entities within the text. However, attention mechanisms and neural network models struggle with syntactic constraints. The quadratic complexity of attention mechanisms also limits their adoption for capturing long-range dependencies between aspect and opinion words in ABSA. This complexity can lead to the misinterpretation of irrelevant contextual words, restricting their effectiveness to short-range dependencies. To address the above problem, we present a novel approach to enhance long-range dependencies between aspect and opinion words in ABSA (MambaForGCN). This approach incorporates syntax-based Graph Convolutional Network (SynGCN) and MambaFormer (Mamba-Transformer) modules to encode input with dependency relations and semantic information. The Multihead Attention (MHA) and Selective State Space model (Mamba) blocks in the MambaFormer module serve as channels to enhance the model with short and long-range dependencies between aspect and opinion words. We also introduce the Kolmogorov-Arnold Networks (KANs) gated fusion, an adaptive feature representation system that integrates SynGCN and MambaFormer and captures non-linear, complex dependencies. Experimental results on three benchmark datasets demonstrate MambaForGCN`s effectiveness, outperforming state-of-the-art (SOTA) baseline models. |
ramirez-orta-etal-2025-rouge | 2,025 | ROUGE-SciQFS: A ROUGE-based Method to Automatically Create Datasets for Scientific Query-Focused Summarization | So far, the task of Scientific Query-Focused Summarization (Sci-QFS) has lagged in development when compared to other areas of Scientific Natural Language Processing because of the lack of data. In this work, we propose a methodology to take advantage of existing collections of academic papers to obtain large-scale datasets for this task automatically. After applying it to the papers from our reading group, we introduce a novel dataset for Sci-QFS composed of 8,695 examples, each one with a query, the sentences of the full text from a paper and the relevance labels for each. After testing several classical and state-of-the-art embedding models on this data, we found that the task of Sci-QFS is far from being solved, although it is relatively straightforward for humans. Surprisingly, we found that classical methods outperformed modern pre-trained Deep Language Models (sometimes by a large margin), showing the need for large datasets to better fine-tune the latter. We share our experiments, data and models at https://github.com/jarobyte91/rouge{\_}sciqfs. |
zhao-etal-2025-commonsense | 2,025 | Commonsense Subgraph for Inductive Relation Reasoning with Meta-learning | In knowledge graphs (KGs), predicting missing relations is a critical reasoning task. Recent subgraph-based models have delved into inductive settings, which aim to predict relations between newly added entities. While these models have demonstrated the ability for inductive reasoning, they only consider the structural information of the subgraph and neglect the loss of semantic information caused by replacing entities with nodes. To address this problem, we propose a novel Commonsense Subgraph Meta-Learning (CSML) model. Specifically, we extract concepts from entities, which can be viewed as high-level semantic information. Unlike previous methods, we use concepts instead of nodes to construct commonsense subgraphs. By combining these with structural subgraphs, we can leverage both structural and semantic information for more comprehensive and rational predictions. Furthermore, we regard concepts as meta-information and employ meta-learning to facilitate rapid knowledge transfer, thus addressing more complex few-shot scenarios. Experimental results confirm the superior performance of our model in both standard and few-shot inductive reasoning. |
zhang-etal-2025-clear | 2,025 | Clear Up Confusion: Iterative Differential Generation for Fine-grained Intent Detection with Contrastive Feedback | Fine-grained intent detection involves identifying a large number of classes with subtle variations. Recently, generating pseudo samples via large language models has attracted increasing attention to alleviate the data scarcity caused by emerging new intents. However, these methods generate samples for each class independently and neglect the relationships between classes, leading to ambiguity in pseudo samples, particularly for fine-grained labels. And, they typically rely on one-time generation and overlook feedback from pseudo samples. In this paper, we propose an iterative differential generation framework with contrastive feedback to generate high-quality pseudo samples and accurately capture the crucial nuances in target class distribution. Specifically, we propose differential guidelines that include potential ambiguous labels to reduce confusion for similar labels. Then we conduct rubric-driven refinement, ensuring the validity and diversity of pseudo samples. Finally, despite one generation, we propose to iteratively generate new samples with contrastive feedback to achieve accurate identification and distillation of target knowledge. Extensive experiments in zero/few-shot and full-shot settings on three datasets verify the effectiveness of our method. |
finch-choi-2025-leveraging | 2,025 | Leveraging Explicit Reasoning for Inference Integration in Commonsense-Augmented Dialogue Models | Open-domain dialogue systems need to grasp social commonsense to understand and respond effectively to human users. Commonsense-augmented dialogue models have been proposed that aim to infer commonsense knowledge from dialogue contexts in order to improve response quality. However, existing approaches to commonsense-augmented dialogue rely on implicit reasoning to integrate commonsense inferences during response generation. In this study, we explore the impact of explicit reasoning against implicit reasoning over commonsense for dialogue response generation. Our findings demonstrate that separating commonsense reasoning into explicit steps for generating, selecting, and integrating commonsense into responses leads to better dialogue interactions, improving naturalness, engagement, specificity, and overall quality. Subsequent analyses of these findings unveil insights into the effectiveness of various types of commonsense in generating responses and the particular response traits enhanced through explicit reasoning for commonsense integration. Our work advances research in open-domain dialogue by achieving a new state-of-the-art in commonsense-augmented response generation. |
wang-etal-2025-integrating | 2,025 | Integrating Group-based Preferences from Coarse to Fine for Cold-start Users Recommendation | Recent studies have demonstrated that cross-domain recommendation (CDR) effectively addresses the cold-start problem. Most approaches rely on transfer functions to generate user representations from the source to the target domain. Although these methods substantially enhance recommendation performance, they exhibit certain limitations, notably the frequent oversight of similarities in user preferences, which can offer critical insights for training transfer functions. Moreover, existing methods typically derive user preferences from historical purchase records or reviews, without considering that preferences operate at three distinct levels: category, brand, and aspect, each influencing decision-making differently. This paper proposes a model that integrates the preferences from coarse to fine levels to improve recommendations for cold-start users. The model leverages historical data from the source domain and external memory networks to generate user representations across different preference levels. A meta-network then transfers these representations to the target domain, where user-item ratings are predicted by aggregating the diverse representations. Experimental results demonstrate that our model outperforms state-of-the-art approaches in addressing the cold-start problem on three CDR tasks. |
mucciaccia-etal-2025-automatic | 2,025 | Automatic Multiple-Choice Question Generation and Evaluation Systems Based on LLM: A Study Case With University Resolutions | Multiple choice questions (MCQs) are often used in both employee selection and training, providing objectivity, efficiency, and scalability. However, their creation is resource-intensive, requiring significant expertise and financial investment. This study leverages large language models (LLMs) and prompt engineering techniques to automate the generation and validation of MCQs, particularly within the context of university regulations. Mainly, two novel approaches are proposed in this work: an automatic question generation system for university resolution and an automatic evaluation system to assess the performance of MCQ generation systems. The generation system combines different prompt engineering techniques and a review process to create well formulated questions. The evaluation system uses prompt engineering combined with an advanced LLM model to assess the integrity of the generated question. Experimental results demonstrate the effectiveness of both systems. The findings highlight the transformative potential of LLMs in educational assessment, reducing the burden on human resources and enabling scalable, cost-effective MCQ generation. |
yu-etal-2025-generating | 2,025 | Generating Commonsense Reasoning Questions with Controllable Complexity through Multi-step Structural Composition | This paper studies the task of generating commonsense reasoning questions (QG) with desired difficulty levels. Compared to traditional shallow questions that can be solved by simple term matching, ours are more challenging. Our answering process requires reasoning over multiple contextual and commonsense clues. That involves advanced comprehension skills, such as abstract semantics learning and missing knowledge inference. Existing work mostly learns to map the given text into questions, lacking a mechanism to control results with the desired complexity. To address this problem, we propose a novel controllable framework. We first derive contextual and commonsense clues involved in reasoning questions from the text. These clues are used to create simple sub-questions. We then aggregate multiple sub-questions to compose complex ones under the guidance of prior reasoning structures. By iterating this process, we can compose a complex QG task based on a series of smaller and simpler QG subtasks. Each subtask serves as a building block for a larger one. Each composition corresponds to an increase in the reasoning step. Moreover, we design a voting verifier to ensure results' validity from multiple views, including answer consistency, reasoning difficulty, and context correlation. Finally, we can learn the optimal QG model to yield thought-provoking results. Evaluations on two typical datasets validate our method. |
li-etal-2025-dna | 2,025 | DnA-Eval: Enhancing Large Language Model Evaluation through Decomposition and Aggregation | The acceleration of Large Language Models (LLMs) research has opened up new possibilities for evaluating generated text. Though LLMs serve as scalable and economical evaluators, how reliable these evaluators is still under-explored. Prior research efforts in the meta-evaluation of LLMs as judges limit the prompting of an LLM to a single use to obtain a final evaluation decision. They then compute the agreement between LLMs' outputs and human labels. This lacks interpretability in understanding the evaluation capability of LLMs. In light of this challenge, we propose DnA-Eval, which breaks down the evaluation process into decomposition and aggregation stages based on pedagogical practices. Our experiments show that it not only provides a more interpretable window for how well LLMs evaluate, but also leads to improvements up to 39.6{\%} for different LLMs on a variety of meta-evaluation benchmarks. |
chu-etal-2025-towards | 2,025 | Towards Faithful Multi-step Reasoning through Fine-Grained Causal-aware Attribution Reasoning Distillation | Despite the remarkable reasoning capabilities demonstrated by large language models (LLM), the substantial computational overhead limits their practices. Some efforts have been directed toward distilling multi-step reasoning capabilities into smaller models through chain-of-thought (CoT). While CoT facilitates multi-step reasoning, the dependencies between reasoning steps are not always clearly discernible, which may lead to inconsistent reasoning. In this paper, we introduce fine-grained attribution reasoning distillation (FARD), which incorporates grounded citations to consolidate the relationships between reasoning steps. Specifically, FARD distills attribution reasoning rationales from LLMs to substitute CoT reasonings, which clarifies the dependencies among reasoning steps. Besides, we regularize the model`s attention pattern by leveraging the causal dependencies between reasoning steps, thereby enhancing the consistency of reasoning. Grounded attribution reasoning also enhances interpretability and verifiability, thereby facilitating faithful reasoning. We evaluate FARD on mathematical and general reasoning benchmarks. The experimental results indicate that FARD outperforms CoT distillation methods in mathematical reasoning, demonstrating its effectiveness. Furthermore, the small models trained with FARD have shown outstanding performance in out-of-distribution reasoning, proving strong generalization capabilities. |
tao-etal-2025-asymkv | 2,025 | AsymKV: Enabling 1-Bit Quantization of KV Cache with Layer-Wise Asymmetric Quantization Configurations | Large language models have shown exceptional capabilities in a wide range of tasks, such as text generation and video generation, among others. However, due to their massive parameter count, these models often require substantial storage space, imposing significant constraints on the machines deploying LLMs. To overcome this limitation, one research direction proposes to compress the models using integer replacements for floating-point numbers, in a process known as Quantization. Some recent studies suggest quantizing the key and value cache (KV Cache) of LLMs, and designing quantization techniques that treat the key and value matrices equivalently. This work delves deeper into the asymmetric structural roles of KV Cache, a phenomenon where the transformer`s output loss is more sensitive to the quantization of key matrices. We conduct a systematic examination of the attention output error resulting from key and value quantization. The phenomenon inspires us to propose an asymmetric quantization strategy. Our approach allows for 1-bit quantization of the KV cache by implementing distinct configurations for key and value matrices. We carry out experiments across a variety of datasets, demonstrating that our proposed model allows for the quantization of up to 75{\%} decoder layers with 1 bit, while simultaneously maintaining performance levels comparable to those of the models with floating parameters. |
zhang-etal-2025-e | 2,025 | E-Bench: Towards Evaluating the Ease-of-Use of Large Language Models | Modern large language models are sensitive to prompts, and another synonymous expression or a typo may lead to unexpected results for the model. Composing an optimal prompt for a specific demand lacks theoretical support and relies entirely on human experimentation, which poses a considerable obstacle to popularizing generative artificial intelligence. However, there is no systematic analysis of the stability of large language models to resist prompt perturbations. In this work, we propose to evaluate the ease-of-use of large language models and construct E-Bench, simulating the actual situation of human use from synonymous perturbation (including paraphrasing, simplification, and colloquialism) and typographical perturbation. Besides we also discuss the combination of these two types of perturbation and analyze the main reasons for performance degradation. Experimental results indicate that with the increase of model size, although the ease-of-use could be significantly improved, there is still a long way to go to build a sufficiently user-friendly model. |
waezi-fani-2025-enhancing | 2,025 | Enhancing Online Grooming Detection via Backtranslation Augmentation | Grooming minors for sexual exploitation become an increasingly significant concern in online conversation platforms. For a safer online experience for minors, machine learning models have been proposed to tap into explicit textual remarks and automate detecting predatory conversations. Such models, however, fall short of real-world applications for the sparse distribution of predatory conversations. In this paper, we propose backtranslation augmentation to augment training datasets with more predatory conversations. Through our experiments on 8 languages from 4 language families using 3 neural translators, we demonstrate that backtranslation augmentation improves models' performance with fewer training epochs for better classification efficacy. Our code and experimental results are available at https://github.com/fani-lab/osprey/tree/coling25. |
feng-etal-2025-causalscore | 2,025 | CausalScore: An Automatic Reference-Free Metric for Assessing Response Relevance in Open-Domain Dialogue Systems | Automatically evaluating the quality of responses in dialogue systems is a challenging yet crucial task. Current metrics often fail to align with human judgments, especially when assessing responses that are grammatically correct. To address this issue, we propose a novel metric, called CausalScore, which assesses the relevance of responses by measuring the causal strength between dialogue histories and responses. The causal strength is estimated by utilizing both unconditional dependence and conditional dependencies from dialogue histories to responses. We compare our metric with the existing competitive metrics in terms of their alignment with human judgements. Our experimental results demonstrate that CausalScore significantly surpasses existing state-of-the-art metrics by aligning better with human judgements. Additionally, we collect a dialogue dataset CGDIALOG+ with human-annotated causal relations and a set of pairwise human judgements to facilitate the development of automatic metrics. |
amidei-etal-2025-exploring | 2,025 | Exploring the Impact of Language Switching on Personality Traits in LLMs | This paper investigates the extent to which LLMs align with humans when personality shifts are associated with language changes. Based on three experiments, that focus on GPT-4o and the Eysenck Personality Questionnaire-Revised (EPQR-A), our initial results reveal a weak yet significant variation in GPT-4o`s personality across languages, indicating that some stem from a language-switching effect rather than translation. Further analysis across five English-speaking countries shows that GPT-4o, leveraging stereotypes, reflects distinct country-specific personality traits. |
wang-etal-2025-llms | 2,025 | LLMs Know What They Need: Leveraging a Missing Information Guided Framework to Empower Retrieval-Augmented Generation | Retrieval-Augmented Generation (RAG) demonstrates great value in alleviating outdated knowledge or hallucination by supplying LLMs with updated and relevant knowledge. However, RAG still faces several challenges in tackling complex multi-hop queries, which require LLMs to perform accurate reasoning and retrieval at each step. Inspired by the human reasoning process, where we progressively search for missing information after acquiring useful clues, it is natural to question whether LLMs have similar capabilities. In this work, we first experimentally verified the ability of LLMs to extract information from the retrieved knowledge as well as to know what is still missing. Based on the above discovery, we propose a Missing Information Guided Retrieve-Extraction-Solving paradigm (MIGRES), where we leverage the identification of missing information to generate a targeted query that steers the subsequent knowledge retrieval. Besides, we design a sentence-level re-ranking filtering approach to filter the irrelevant content from the document, along with the information extraction capability of LLMs to extract useful information from denoised documents. Extensive experiments conducted on multiple public datasets reveal the superiority of the proposed MIGRES method, and analytical experiments demonstrate the effectiveness of our proposed modules. Code and data are released in https://github.com/AdelWang/MIGRES. |
wei-etal-2025-chain | 2,025 | Chain-of-Specificity: Enhancing Task-Specific Constraint Adherence in Large Language Models | Large Language Models (LLMs) exhibit remarkable generative capabilities, enabling the generation of valuable information. Despite these advancements, previous research found that LLMs sometimes struggle with adhering to specific constraints, such as being in a specific place or at a specific time, and at times even overlook them, which leads to responses that are either too generic or not fully satisfactory. Existing approaches attempted to address this issue by decomposing and rewriting input instructions or reflecting on prior failings, yet they fall short in adequately emphasizing specific constraints and unlocking the underlying knowledge, such as programming within the context of software development. In response, this paper proposes a simple yet effective method called Chain-of-Specificity (CoS). Specifically, CoS emphasizes the specific constraints in the input instructions, unlocks knowledge within LLMs, and refines responses. Experiments conducted on publicly available and self-built complex datasets demonstrate that CoS outperforms existing methods in enhancing generated content, especially in terms of specificity. Additionally, as the number of specific constraints increases, other baselines falter, while CoS still performs well. Moreover, we show that distilling responses generated by CoS effectively enhances the ability of smaller models to follow constrained instructions. |
liu-etal-2025-transliterations | 2,025 | How Transliterations Improve Crosslingual Alignment | Recent studies have shown that post-aligning multilingual pretrained language models (mPLMs) using alignment objectives on both original and transliterated data can improve crosslingual alignment. This improvement further leads to better crosslingual transfer performance. However, it remains unclear how and why a better crosslingual alignment is achieved, as this technique only involves transliterations, and does not use any parallel data. This paper attempts to explicitly evaluate the crosslingual alignment and identify the key elements in transliteration-based approaches that contribute to better performance. For this, we train multiple models under varying setups for two pairs of related languages: (1) Polish and Ukrainian and (2) Hindi and Urdu. To assess alignment, we define four types of similarities based on sentence representations. Our experimental results show that adding transliterations alone improves the overall similarities, even for random sentence pairs. With the help of auxiliary transliteration-based alignment objectives, especially the contrastive objective, the model learns to distinguish matched from random pairs, leading to better crosslingual alignment. However, we also show that better alignment does not always yield better downstream performance, suggesting that further research is needed to clarify the connection between alignment and performance. The code implementation is based on \url{https://github.com/cisnlp/Transliteration-PPA}. |
wang-etal-2025-gl | 2,025 | GL-GAN: Perceiving and Integrating Global and Local Styles for Handwritten Text Generation with Mamba | Handwritten text generation (HTG) aims to synthesize handwritten samples by imitating a specific writer, which has a wide range of applications and thus has significant research value. However, current studies on HTG are confronted with a main bottleneck: dominant models lack the ability to perceive and integrate handwriting styles, which affects the realism of the synthesized samples. In this paper, we propose GL-GAN, which effectively captures and integrates global and local styles. Specifically, we propose a Hybrid Style Encoder (HSE) that combines a state space model (SSM) and convolution to capture multilevel style features through various receptive fields. The captured style features are then fed to the proposed Dynamic Feature Enhancement Module (DFEM), which integrates these features by adaptively modeling the entangled relationships between multilevel styles and removing redundant details. Extensive experiments on two widely used handwriting datasets demonstrate that our GL-GAN is an effective HTG model and outperforms state-of-the-art models remarkably. Our code is publicly available at:https://github.com/Fyzjym/GL-GAN. |
tilli-vu-2025-discrete | 2,025 | Discrete Subgraph Sampling for Interpretable Graph based Visual Question Answering | Explainable artificial intelligence (XAI) aims to make machine learning models more transparent. While many approaches focus on generating explanations post-hoc, interpretable approaches, which generate the explanations intrinsically alongside the predictions, are relatively rare. In this work, we integrate different discrete subset sampling methods into a graph-based visual question answering system to compare their effectiveness in generating interpretable explanatory subgraphs intrinsically. We evaluate the methods on the dataset and show that the integrated methods effectively mitigate the performance trade-off between interpretability and answer accuracy, while also achieving strong co-occurrences between answer and question tokens. Furthermore, we conduct a human evaluation to assess the interpretability of the generated subgraphs using a comparative setting with the extended Bradley-Terry model, showing that the answer and question token co-occurrence metrics strongly correlate with human preferences. Our source code is publicly available. |
lynn-etal-2025-multiple | 2,025 | From Multiple-Choice to Extractive QA: A Case Study for English and Arabic | The rapid evolution of Natural Language Processing (NLP) has favoured major languages such as English, leaving a significant gap for many others due to limited resources. This is especially evident in the context of data annotation, a task whose importance cannot be underestimated, but which is time-consuming and costly. Thus, any dataset for resource-poor languages is precious, in particular when it is task-specific. Here, we explore the feasibility of repurposing an existing multilingual dataset for a new NLP task: we repurpose a subset of the BELEBELE dataset (Bandarkar et al., 2023), which was designed for multiple-choice question answering (MCQA), to enable the more practical task of extractive QA (EQA) in the style of machine reading comprehension. We present annotation guidelines and a parallel EQA dataset for English and Modern Standard Arabic (MSA). We also present QA evaluation results for several monolingual and cross-lingual QA pairs including English, MSA, and five Arabic dialects. We aim to help others adapt our approach for the remaining 120 BELEBELE language variants, many of which are deemed under-resourced. We also provide a thorough analysis and share insights to deepen understanding of the challenges and opportunities in NLP task reformulation. |
peng-zhang-2025-enhancing | 2,025 | Enhancing Knowledge Distillation of Large Language Models through Efficient Multi-Modal Distribution Alignment | Knowledge distillation (KD) is an effective model compression method that can transfer the internal capabilities of large language models (LLMs) to smaller ones. However, the multi-modal probability distribution predicted by teacher LLMs causes difficulties for student models to learn. In this paper, we first demonstrate the importance of multi-modal distribution alignment with experiments and then highlight the inefficiency of existing KD approaches in learning multi-modal distributions. To address this problem, we propose Ranking Loss based Knowledge Distillation (RLKD), which encourages the consistency of the ranking of peak predictions between the teacher and student models. By incorporating word-level ranking loss, we ensure excellent compatibility with existing distillation objectives while fully leveraging the fine-grained information between different categories in peaks of two predicted distribution. Experimental results demonstrate that our method enables the student model to better learn the multi-modal distributions of the teacher model, leading to a significant performance improvement in various downstream tasks. |
he-etal-2025-dialoguemmt | 2,025 | DialogueMMT: Dialogue Scenes Understanding Enhanced Multi-modal Multi-task Tuning for Emotion Recognition in Conversations | Emotion recognition in conversations (ERC) has garnered significant attention from the research community. However, due to the complexity of visual scenes and dialogue contextual dependencies in conversations, previous ERC methods fail to handle emotional cues from both visual sources and discourse structures. Furthermore, existing state-of-the-art ERC models are trained and tested separately on each single ERC dataset, not verifying their effectiveness across multiple datasets simultaneously. To address these challenges, this paper proposes an innovative framework for ERC, called Dialogue Scenes Understanding Enhanced Multi-modal Multi-task Tuning (DialogueMMT). More concretely, a novel video-language connector is applied within the large vision-language model for capturing video features effectively. Additionally, we utilize multi-task instruction tuning with a unified ERC dataset to enhance the model`s understanding of multi-modal dialogue scenes and employ a chain-of-thought strategy to improve emotion classification performance. Extensive experimental results on three benchmark ERC datasets indicate that the proposed DialogueMMT framework consistently outperforms existing state-of-the-art approaches in terms of overall performance. |
zhai-etal-2025-learning | 2,025 | Learning Transition Patterns by Large Language Models for Sequential Recommendation | Large Language Models (LLMs) have demonstrated powerful performance in sequential recommendation due to their robust language modeling and comprehension capabilities. In such paradigms, the item texts of interaction sequences are formulated as sentences and LLMs are utilized to learn language representations or directly generate target item texts by incorporating instructions. Despite their promise, these methods solely focus on modeling the mapping from sequential texts to target items, neglecting the relationship between the items in an interaction sequence. This results in a failure to learn the transition patterns between items, which reflect the dynamic change in user preferences and are crucial for predicting the next item. To tackle this issue, we propose a novel framework for mapping the sequential item texts to the sequential item IDs, named \textbf{ST2SI}. Specifically, we first introduce multi-query input and item linear projection (ILP) to model the conditional probability distribution of items. Then, we further propose ID alignment to address misalignment between item texts and item IDs by instruction tuning. Finally, we propose efficient ILP tuning to adapt flexibly to different scenarios, requiring only training a linear layer to achieve competitive performance. Extensive experiments on six real-world datasets show our approach outperforms the best baselines by 7.33{\%} in NDCG@10, 4.65{\%} in Recall@10, and 8.42{\%} in MRR. |
do-etal-2025-aligning | 2,025 | Aligning Large Language Models with Human Opinions through Persona Selection and Value--Belief--Norm Reasoning | Reasoning and predicting human opinions with large language models (LLMs) is essential yet challenging. Current methods employ role-playing with personae but face two major issues: LLMs are sensitive to even a single irrelevant persona, skewing predictions by up to 30{\%}; and LLMs fail to reason strategically over personae. We propose Chain-of-Opinion (COO), a simple four-step solution modeling which and how to reason with personae, inspired by the Value{--}Belief{--}Norm (VBN) theory. COO differentiates between explicit personae (demographics and ideology) and implicit personae (historical opinions), involves: (1) filtering irrelevant attributes from explicit personae; (2) ranking implicit personae into a preferential list for selecting top-k; (3) applying novel VBN reasoning to extract user environmental and personal value, belief, and norm variables for accurate and reliable predictions; and (4) iterating VBN reasoning with progressively larger lists of implicit personae to handle potential persona insufficiency. COO efficiently achieves new state-of-the-art opinion prediction via prompting with only 5 inference calls, improving prior techniques by up to 4{\%}. Notably, fine-tuning LMs with COO`s data results in significantly better opinion-aligned models, by up to 23{\%}. |
li-etal-2025-mimotable | 2,025 | MiMoTable: A Multi-scale Spreadsheet Benchmark with Meta Operations for Table Reasoning | Extensive research has been conducted to explore the capability of Large Language Models (LLMs) for table reasoning and has significantly improved the performance on existing benchmarks. However, tables and user questions in real-world applications are more complex and diverse, presenting an unignorable gap compared to the existing benchmarks. To fill the gap, we propose a \textbf{M}ult\textbf{i}-scale spreadsheet benchmark with \textbf{M}eta \textbf{o}perations for \textbf{Table} reasoning, named as MiMoTable. Specifically, MiMoTable incorporates two key features. First, the tables in MiMoTable are all spreadsheets used in real-world scenarios, which cover seven domains and contain different types. Second, we define a new criterion with six categories of meta operations for measuring the difficulty of each question in MiMoTable, simultaneously as a new perspective for measuring the difficulty of the existing benchmarks. Experimental results show that Claude-3.5-Sonnet achieves the best performance with 77.4{\%} accuracy, indicating that there is still significant room to improve for LLMs on MiMoTable. Furthermore, we grade the difficulty of existing benchmarks according to our new criteria. Experiments have shown that the performance of LLMs decreases as the difficulty of benchmarks increases, thereby proving the effectiveness of our proposed new criterion. |
saeed-etal-2025-implicit | 2,025 | Implicit Discourse Relation Classification For Nigerian Pidgin | Nigerian Pidgin (NP) is an English-based creole language spoken by nearly 100 million people across Nigeria, and is still low-resource in NLP. In particular, there are currently no available discourse parsing tools, which, if available, would have the potential to improve various downstream tasks. Our research focuses on implicit discourse relation classification (IDRC) for NP, a task which, even in English, is not easily solved by prompting LLMs, but requires supervised training. {\%} With this in mind, we have developed a framework for the task, which could also be used by researchers for other English-lexified languages. We systematically compare different approaches to the low resource IDRC task: in one approach, we use English IDRC tools directly on the NP text as well as on their English translations (followed by a back-projection of labels). In another approach, we create a synthetic discourse corpus for NP, in which we automatically translate the English discourse-annotated corpus PDTB to NP, project PDTB labels, and then train an NP IDR classifier. The latter approach of training a {\textquotedblleft}native{\textquotedblright} NP classifier outperforms our baseline by 13.27{\%} and 33.98{\%} in f$_{1}$ score for 4-way and 11-way classification, respectively. |
ji-chen-2025-many | 2,025 | How Many Languages Make Good Multilingual Instruction Tuning? A Case Study on BLOOM | Instruction tuning a large language model with multiple languages can prepare it for multilingual downstream tasks. Nonetheless, it is yet to be determined whether having a handful of languages is sufficient, or whether the benefits increase with the inclusion of more. By fine-tuning large multilingual models on 1 to 52 languages, we present a case study on BLOOM to understand three pertinent factors affecting performance: the number of languages, language exposure, and similarity between training and test languages. Overall we found that 1) expanding language coverage in multilingual instruction tuning proves to be beneficial; 2) accuracy often significantly boots if the test language appears in the instruction mixture; 3) languages' genetic features correlate with cross-lingual transfer more than merely the number of language but different languages benefit to various degrees. |
gao-etal-2025-gradient | 2,025 | Gradient Inversion Attack in Federated Learning: Exposing Text Data through Discrete Optimization | Federated learning has emerged as a potential solution to overcome the bottleneck posed by the near exhaustion of public text data in training large language models. There are claims that the strategy of exchanging gradients allows using text data including private information. Although recent studies demonstrate that data can be reconstructed from gradients, the threat for text data seems relatively small due to its sensitivity to even a few token errors. However, we propose a novel attack method FET, indicating that it is possible to Fully Expose Text data from gradients. Unlike previous methods that optimize continuous embedding vectors, we directly search for a text sequence with gradients that match the known gradients. First, we infer the total number of tokens and the unique tokens in the target text data from the gradients of the embedding layer. Then we develop a discrete optimization algorithm, which globally explores the solution space and precisely refines the obtained solution, incorporating both global and local search strategies. We also find that gradients of the fully connected layer are dominant, providing sufficient guidance for the optimization process. Our experiments show a significant improvement in attack performance, with an average increase of 39{\%} for TinyBERT-6, 20{\%} for BERT-base and 15{\%} for BERT-large in exact match rates across three datasets. These findings highlight serious privacy risks in text data, suggesting that using smaller models is not an effective privacy-preserving strategy. |
wang-etal-2025-simulating | 2,025 | Simulating Dual-Process Thinking in Dialogue Topic Shift Detection | Previous work on dialogue topic shift detection has primarily focused on shallow local reasoning, overlooking the importance of considering the global historical structure and local details to elucidate the underlying causes of topic shift. To address the above two issues, we introduce the dual-process theory to this task and design a novel Dual-Module Framework DMF (i.e., intuition and reasoning module) for dialogue topic shift detection to emulate this cognitive process. Specifically, the intuition module employs Large Language Models (LLMs) to extract and store the global topic structure of historical dialogue, while the reasoning module introduces a LLM to generate reasoning samples between the response and the most recent topic of historical dialogue, thereby providing local detail explanations for topic shift. Moreover, we distill the dual-module framework into a small generative model to facilitate more precise reasoning. The experimental results on three public datasets show that our DMF outperforms the state-of-the-art baselines. |
sun-etal-2025-compliance | 2,025 | A Compliance Checking Framework Based on Retrieval Augmented Generation | The text-based compliance checking aims to verify whether a company`s business processes comply with laws, regulations, and industry standards using NLP techniques. Existing methods can be divided into two categories: Logic-based methods offer the advantage of precise and reliable reasoning processes but lack flexibility. Semantic embedding methods are more generalizable; however, they may lose structured information and lack logical coherence. To combine the strengths of both approaches, we propose a compliance checking framework based on Retrieval-Augmented Generation (RAG). This framework includes a static layer for storing factual knowledge, a dynamic layer for storing regulatory and business process information, and a computational layer for retrieval and reasoning. We employ an eventic graph to structurally describe regulatory information as we recognize that the knowledge in regulatory documents is centered not on entities but on actions and states. We conducted experiments on Chinese and English compliance checking datasets. The results demonstrate that our framework consistently achieves state-of-the-art results across various scenarios, surpassing other baselines. |
yin-etal-2025-midlm | 2,025 | MIDLM: Multi-Intent Detection with Bidirectional Large Language Models | Decoder-only Large Language Models (LLMs) have demonstrated exceptional performance in language generation, exhibiting broad capabilities across various tasks. However, the application to label-sensitive language understanding tasks remains challenging due to the limitations of their autoregressive architecture, which restricts the sharing of token information within a sentence. In this paper, we address the Multi-Intent Detection (MID) task and introduce MIDLM, a bidirectional LLM framework that incorporates intent number detection and multi-intent selection. This framework allows autoregressive LLMs to leverage bidirectional information awareness through post-training, eliminating the need for training the models from scratch. Comprehensive evaluations across 8 datasets show that MIDLM consistently outperforms both existing vanilla models and pretrained baselines, demonstrating its superior performance in the MID task. |
song-etal-2025-prosparse | 2,025 | ProSparse: Introducing and Enhancing Intrinsic Activation Sparsity within Large Language Models | Activation sparsity refers to the existence of considerable weakly-contributed elements among activation outputs, serving as a promising paradigm for accelerating model inference. Nevertheless, most large language models (LLMs) adopt activation functions without intrinsic activation sparsity (e.g., GELU and Swish). Some recent efforts have explored introducing ReLU or its variants as the substitutive activation function to pursue activation sparsity and acceleration, but few can simultaneously obtain high activation sparsity and comparable model performance. This paper introduces a simple and effective method named {\textquotedblleft}ProSparse{\textquotedblright} to sparsify LLMs while achieving both targets. Specifically, after introducing ReLU activation, ProSparse adopts progressive sparsity regularization with a factor smoothly increasing for multiple stages. This can enhance activation sparsity and mitigate performance degradation by avoiding radical shifts in activation distributions. With ProSparse, we obtain high sparsity of 89.32{\%} for LLaMA2-7B, 88.80{\%} for LLaMA2-13B, and 87.89{\%} for end-size MiniCPM-1B, respectively, with comparable performance to their original Swish-activated versions. These present the most sparsely activated models among open-source LLaMA versions and competitive end-size models. Inference acceleration experiments further demonstrate the significant practical acceleration potential of LLMs with higher activation sparsity, obtaining up to 4.52x inference speedup. |
tang-etal-2025-reasoning | 2,025 | Reasoning-Oriented and Analogy-Based Methods for Locating and Editing in Zero-Shot Event-Relational Reasoning | Zero-shot event-relational reasoning is an important task in natural language processing, and existing methods jointly learn a variety of event-relational prefixes and inference-form prefixes to achieve such tasks. However, training prefixes consumes large computational resources and lacks interpretability. Additionally, learning various relational and inferential knowledge inefficiently exploits the connections between tasks. Therefore, we first propose a method for Reasoning-Oriented Locating and Editing (ROLE), which locates and edits the key modules of the language model for reasoning about event relations, enhancing interpretability and also resource-efficiently optimizing the reasoning ability. Subsequently, we propose a method for Analogy-Based Locating and Editing (ABLE), which efficiently exploits the similarities and differences between tasks to optimize the zero-shot reasoning capability. Experimental results show that ROLE improves interpretability and reasoning performance with reduced computational cost. ABLE achieves SOTA results in zero-shot reasoning. |
sahu-etal-2025-leveraging | 2,025 | Leveraging Language Models for Summarizing Mental State Examinations: A Comprehensive Evaluation and Dataset Release | Mental health disorders affect a significant portion of the global population, with diagnoses primarily conducted through Mental State Examinations (MSEs). MSEs serve as structured assessments to evaluate behavioral and cognitive functioning across various domains, aiding mental health professionals in diagnosis and treatment monitoring. However, in developing countries, access to mental health support is limited, leading to an overwhelming demand for mental health professionals. Resident doctors often conduct initial patient assessments and create summaries for senior doctors, but their availability is constrained, resulting in extended patient wait times. This study addresses the challenge of generating concise summaries from MSEs through the evaluation of various language models. Given the scarcity of relevant mental health conversation datasets, we developed a 12-item descriptive MSE questionnaire and collected responses from 405 participants, resulting in 9720 utterances covering diverse mental health aspects. Subsequently, we assessed the performance of five well-known pre-trained summarization models, both with and without fine-tuning, for summarizing MSEs. Our comprehensive evaluation, leveraging metrics such as ROUGE, SummaC, and human evaluation, demonstrates that language models can generate automated coherent MSE summaries for doctors. With this paper, we release our collected conversational dataset and trained models publicly for the mental health research community. |
gralinski-etal-2025-oddballness | 2,025 | Oddballness: universal anomaly detection with language models | We present a new method to detect anomalies in texts (in general: in sequences of any data), using language models, in a totally unsupervised manner. The method considers probabilities (likelihoods) generated by a language model, but instead of focusing on low-likelihood tokens, it considers a new metric defined in this paper: oddballness. Oddballness measures how {\textquotedblleft}strange{\textquotedblright} a given token is according to the language model. We demonstrate in grammatical error detection tasks (a specific case of text anomaly detection) that oddballness is better than just considering low-likelihood events, if a totally unsupervised setup is assumed. |
li-etal-2025-cmmath | 2,025 | CMMaTH: A Chinese Multi-modal Math Skill Evaluation Benchmark for Foundation Models | With the rapid advancements in multimodal large language models, evaluating their multimodal mathematical capabilities continues to receive wide attention. Although datasets such as MathVista have been introduced for evaluating mathematical capabilities in multimodal scenarios, there remains a lack of evaluation tools and datasets tailored for fine-grained assessment in Chinese K12 education. To systematically evaluate the ability of multimodal large models to solve Chinese multimodal mathematical problems, we propose a Chinese Multi-modal Math Skill Evaluation Benchmark (CMMaTH), containing 23,856 multimodal K12 math related questions, making it the largest Chinese multimodal mathematical problem benchmark to date. CMMaTH includes questions ranging from elementary to high school levels, offering greater diversity in problem types, solution goals, visual elements, detailed knowledge points, and standard solution annotations. To facilitate stable, fast, and cost-free model evaluation, we have developed an open-source tool called GradeGPT, which is integrated with the CMMaTH dataset. Our data and code are available at \url{https://github.com/zzli2022/CMMaTH}. |
gao-etal-2025-efficient | 2,025 | Efficient Tool Use with Chain-of-Abstraction Reasoning | To achieve faithful reasoning that aligns with human expectations, large language models (LLMs) need to ground their reasoning to real-world knowledge (e.g., web facts, math and physical rules). Tools help LLMs access this external knowledge, but there remains challenges for fine-tuning LLM agents (e.g., Toolformer) to invoke tools in multi-step reasoning problems, where inter-connected tool calls require holistic and efficient tool usage planning. In this work, we propose a new method for LLMs to better leverage tools in multi-step reasoning. Our method, Chain-of-Abstraction (CoA), trains LLMs to first decode reasoning chains with abstract placeholders, and then call domain tools to reify each reasoning chain by filling in specific knowledge. This planning with abstract chains enables LLMs to learn more general reasoning strategies, which are robust to shifts of domain knowledge (e.g., math results) relevant to different reasoning questions. It also allows LLMs to perform decoding and calling of external tools in parallel, which avoids the inference delay caused by waiting for tool responses. In mathematical reasoning and Wiki QA domains, we show that our method consistently outperforms previous chain-of-thought and tool-augmented baselines on both in-distribution and out-of-distribution test sets, with an average {\textasciitilde}6{\%} absolute QA accuracy improvement. LLM agents trained with our method also show more efficient tool use, with inference speed being on average {\textasciitilde}1.4x faster than baseline tool-augmented LLMs. |
mohamed-al-azani-2025-enhancing | 2,025 | Enhancing Arabic NLP Tasks through Character-Level Models and Data Augmentation | This study introduces a character-level approach specifically designed for Arabic NLP tasks, offering a novel and highly effective solution to the unique challenges inherent in Arabic language processing. It presents a thorough comparative study of various character-level models, including Convolutional Neural Networks (CNNs), pre-trained transformers (CANINE), and Bidirectional Long Short-Term Memory networks (BiLSTMs), assessing their performance and exploring the impact of different data augmentation techniques on enhancing their effectiveness. Additionally, it introduces two innovative Arabic-specific data augmentation methods{---}vowel deletion and style transfer{---}and rigorously evaluates their effectiveness. The proposed approach was evaluated on Arabic privacy policy classification task as a case study, demonstrating significant improvements in model performance, reporting a micro-averaged F1-score of 93.8{\%}, surpassing state-of-the-art models. |
kaneko-etal-2025-gaps | 2,025 | The Gaps between Fine Tuning and In-context Learning in Bias Evaluation and Debiasing | The output tendencies of PLMs vary markedly before and after FT due to the updates to the model parameters. These divergences in output tendencies result in a gap in the social biases of PLMs. For example, there exits a low correlation between intrinsic bias scores of a PLM and its extrinsic bias scores under FT-based debiasing methods. Additionally, applying FT-based debiasing methods to a PLM leads to a decline in performance in downstream tasks. On the other hand, PLMs trained on large datasets can learn without parameter updates via ICL using prompts. ICL induces smaller changes to PLMs compared to FT-based debiasing methods. Therefore, we hypothesize that the gap observed in pre-trained and FT models does not hold true for debiasing methods that use ICL. In this study, we demonstrate that ICL-based debiasing methods show a higher correlation between intrinsic and extrinsic bias scores compared to FT-based methods. Moreover, the performance degradation due to debiasing is also lower in the ICL case compared to that in the FT case. |
zhang-etal-2025-llm | 2,025 | LLM Sensitivity Challenges in Abusive Language Detection: Instruction-Tuned vs. Human Feedback | The capacity of large language models (LLMs) to understand and distinguish socially unacceptable texts enables them to play a promising role in abusive language detection. However, various factors can affect their sensitivity. In this work, we test whether LLMs have an unintended bias in abusive language detection, i.e., whether they predict more or less of a given abusive class than expected in zero-shot settings. Our results show that instruction-tuned LLMs tend to under-predict positive classes, since datasets used for tuning are dominated by the negative class. On the contrary, models fine-tuned with human feedback tend to be overly sensitive. In an exploratory approach to mitigate these issues, we show that label frequency in the prompt helps with the significant over-prediction. |
gu-etal-2025-improving | 2,025 | Improving Automatic Grammatical Error Annotation for Chinese Through Linguistically-Informed Error Typology | Comprehensive error annotation is essential for developing effective Grammatical Error Correction (GEC) systems and delivering meaningful feedback to learners. This paper introduces improvements to automatic grammatical error annotation for Chinese. Our refined framework addresses language-specific challenges that cause common spelling errors in Chinese, including pronunciation similarity, visual shape similarity, specialized participles, and word ordering. In a case study, we demonstrated our system`s ability to provide detailed feedback on 12-16{\%} of all errors by identifying them under our new error typology, specific enough to uncover subtle differences in error patterns between L1 and L2 writings. In addition to improving automated feedback for writers, this work also highlights the value of incorporating language-specific features in NLP systems. |
shirafuji-etal-2025-bias | 2,025 | Bias Vector: Mitigating Biases in Language Models with Task Arithmetic Approach | The use of language models (LMs) has increased considerably in recent years, and the biases and stereotypes in training data that are reflected in the LM outputs are causing social problems. In this paper, inspired by the task arithmetic, we propose the {\textquotedblleft}Bias Vector{\textquotedblright} method for the mitigation of these LM biases. The Bias Vector method does not require manually created debiasing data. The three main steps of our approach involve: (1) continual training the pre-trained LMs on biased data using masked language modeling; (2) constructing the Bias Vector as the difference between the weights of the biased LMs and those of pre-trained LMs; and (3) subtracting the Bias Vector from the weights of the pre-trained LMs for debiasing. We evaluated the Bias Vector method on the SEAT across three LMs and confirmed an average improvement of 0.177 points. We demonstrated that the Bias Vector method does not degrade the LM performance on downstream tasks in the GLUE benchmark. In addition, we examined the impact of scaling factors, which control the magnitudes of Bias Vectors, with effect sizes on the SEAT and conducted a comprehensive evaluation of our debiased LMs across both the SEAT and GLUE benchmarks. |
li-etal-2025-topology | 2,025 | Topology-of-Question-Decomposition: Enhancing Large Language Models with Information Retrieval for Knowledge-Intensive Tasks | Large language models (LLMs) are increasingly deployed for general problem-solving across various domains yet remain constrained to chaining immediate reasoning steps and depending solely on parametric knowledge. Integrating an information retrieval system directly into the reasoning process of LLMs can improve answer accuracy but might disrupt the natural reasoning sequence. Consequently, LLMs may underperform in complex, knowledge-intensive tasks requiring multiple reasoning steps, extensive real-world knowledge, or critical initial decisions. To overcome these challenges, we introduce a novel framework, Topology-of-Question-Decomposition (ToQD), which activates retrieval only when necessary. Globally, ToQD guides LLMs in constructing a topology graph from the input question, each node representing a sub-question. Locally, ToQD employs self-verify inference to determine whether a sub-question should retrieve relevant documents, necessitate further decomposition, or directly provide an answer. Experiments demonstrate that ToQD achieves superior performance and robustness in complex, knowledge-intensive tasks, significantly enhancing system response efficiency. |
li-li-2025-hne | 2,025 | t-HNE: A Text-guided Hierarchical Noise Eliminator for Multimodal Sentiment Analysis | In the Multimodal Sentiment Analysis task, most existing approaches focus on extracting modality-consistent information from raw unimodal data and integrating it into multimodal representations for sentiment classification. However, these methods often assume that all modalities contribute equally to model performance, prioritizing the extraction and enhancement of consistent information, while overlooking the adverse effects of noise caused by modality inconsistency. In contrast to these approaches, this paper introduces a novel approach namely text-guided Hierarchical Noise Eliminator (t-HNE). This model consists of a two-stage denoising phase and a feature recovery phase. Firstly, textual information is injected into both visual and acoustic modalities using an attention mechanism, aiming to reduce intra-modality noise in the visual and acoustic representations. Secondly, it further mitigates inter-modality noise by maximizing the mutual information between textual representations and the respective visual and acoustic representations. Finally, to address the potential loss of modality-invariant information during denoising, the fused multimodal representation is refined through contrastive learning with each unimodal representation except the textual. Extensive experiments conducted on the CMU-MOSI and CMU-MOSEI datasets demonstrate the efficacy of our approach. |
mao-etal-2025-alympics | 2,025 | ALYMPICS: LLM Agents Meet Game Theory | Game theory is a branch of mathematics that studies strategic interactions among rational agents. We propose Alympics (Olympics for Agents), a systematic framework utilizing Large Language Model (LLM) agents for empirical game theory research. Alympics creates a versatile platform for studying complex game theory problems, bridging the gap between theoretical game theory and empirical investigations by providing a controlled environment for simulating human-like strategic interactions with LLM agents. In our pilot case study, the {\textquotedblleft}Water Allocation Challenge{\textquotedblright}, we explore Alympics through a challenging strategic game focused on the multi-round auction of scarce survival resources. This study demonstrates the framework`s ability to qualitatively and quantitatively analyze game determinants, strategies, and outcomes. Additionally, we conduct a comprehensive human assessment and an in-depth evaluation of LLM agents in rational strategic decision-making scenarios. Our findings highlight LLM agents' potential to advance game theory knowledge and expand the understanding of their proficiency in emulating human strategic behavior. |
huang-etal-2025-towards | 2,025 | Towards Adaptive Mechanism Activation in Language Agent | Language Agent could be endowed with different mechanisms for autonomous task accomplishment. Current agents typically rely on a fixed mechanism or a set of mechanisms activated in a predefined order, limiting their adaptation to varied potential task solution structures. To this end, this paper proposes Adaptive Language Agent Mechanism Activation Learning with Self-Exploration (ALAMA), which focuses on optimizing mechanism activation adaptability without reliance on expert models. Initially, it builds a harmonized agent framework (UniAct) to Unify different mechanisms via Actions. Then it leverages a training-efficient optimization method based on self-exploration to enable the UniAct to adaptively activate the appropriate mechanisms according to the potential characteristics of the task. Experimental results demonstrate significant improvements in downstream agent tasks, affirming the effectiveness of our approach in facilitating more dynamic and context-sensitive mechanism activation. |
lei-etal-2025-scaffolding | 2,025 | Scaffolding Coordinates to Promote Vision-Language Coordination in Large Multi-Modal Models | State-of-the-art Large Multi-Modal Models (LMMs) have demonstrated exceptional capabilities in vision-language tasks. Despite their advanced functionalities, the performances of LMMs are still limited in challenging scenarios that require complex reasoning with multiple levels of visual information. Existing prompting techniques for LMMs focus on either improving textual reasoning or leveraging tools for image preprocessing, lacking a simple and general visual prompting scheme to promote vision-language coordination in LMMs. In this work, we propose SCAFFOLD prompting that scaffolds coordinates to promote vision-language coordination. Specifically, SCAFFOLD overlays a dot matrix within the image as visual information anchors and leverages multi-dimensional coordinates as textual positional references. Extensive experiments on a wide range of challenging vision-language tasks demonstrate the superiority of SCAFFOLD over the textual Chain-of-Thought prompting. |
xie-etal-2025-retrieval | 2,025 | Retrieval Augmented Instruction Tuning for Open NER with Large Language Models | The strong capability of large language models (LLMs) has been applied to information extraction (IE) through either retrieval augmented prompting or instruction tuning (IT). However, the best way to incorporate information with LLMs for IE remains an open question. In this paper, we explore Retrieval Augmented Instruction Tuning (RA-IT) for IE, focusing on the task of open named entity recognition (NER). Specifically, for each training sample, we retrieve semantically similar examples from the training dataset as the context and prepend them to the input of the original instruction. To evaluate our RA-IT approach more thoroughly, we construct a Chinese IT dataset for open NER and evaluate RA-IT in both English and Chinese scenarios. Experimental results verify the effectiveness of RA-IT across various data sizes and in both English and Chinese scenarios. We also conduct thorough studies to explore the impacts of various retrieval strategies in the proposed RA-IT framework. |
lin-etal-2025-rethinking | 2,025 | Rethinking Vocabulary Augmentation: Addressing the Challenges of Low-Resource Languages in Multilingual Models | The performance of multilingual language models (MLLMs) is notably inferior for low-resource languages (LRL) compared to high-resource ones, primarily due to the limited available corpus during the pre-training phase. This inadequacy stems from the under-representation of low-resource language words in the subword vocabularies of MLLMs, leading to their misidentification as unknown or incorrectly concatenated subwords. Previous approaches are based on frequency sorting to select words for augmenting vocabularies. However, these methods overlook the fundamental disparities between model representation distributions and frequency distributions. To address this gap, we introduce a novel Entropy-Consistency Word Selection (ECWS) method, which integrates semantic and frequency metrics for vocabulary augmentation. Our results indicate an improvement in performance, supporting our approach as a viable means to enrich vocabularies inadequately represented in current MLLMs. |
du-etal-2025-hawkes | 2,025 | Hawkes based Representation Learning for Reasoning over Scale-free Community-structured Temporal Knowledge Graphs | Temporal knowledge graph (TKG) reasoning has become a hot topic due to its great value in many practical tasks. The key to TKG reasoning is modeling the structural information and evolutional patterns of the TKGs. While great efforts have been devoted to TKG reasoning, the structural and evolutional characteristics of real-world networks have not been considered. In the aspect of structure, real-world networks usually exhibit clear community structure and scale-free (long-tailed distribution) properties. In the aspect of evolution, the impact of an event decays with the time elapsing. In this paper, we propose a novel TKG reasoning model called Hawkes process-based Evolutional Representation Learning Network (HERLN), which learns structural information and evolutional patterns of a TKG simultaneously, considering the characteristics of real-world networks: community structure, scale-free and temporal decaying. First, we find communities in the input TKG to make the encoding get more similar intra-community embeddings. Second, we design a Hawkes process-based relational graph convolutional network to cope with the event impact-decaying phenomenon. Third, we design a conditional decoding method to alleviate biases towards frequent entities caused by long-tailed distribution. Experimental results show that HERLN achieves significant improvements over the state-of-the-art models. |
zhang-etal-2025-intention | 2,025 | Intention Analysis Makes LLMs A Good Jailbreak Defender | Aligning large language models (LLMs) with human values, particularly when facing complex and stealthy jailbreak attacks, presents a formidable challenge. Unfortunately, existing methods often overlook this intrinsic nature of jailbreaks, which limits their effectiveness in such complex scenarios. In this study, we present a simple yet highly effective defense strategy, i.e., Intention Analysis (IA). IA works by triggering LLMs' inherent self-correct and improve ability through a two-stage process: 1) analyzing the essential intention of the user input, and 2) providing final policy-aligned responses based on the first round conversation. Notably,IA is an inference-only method, thus could enhance LLM safety without compromising their helpfulness. Extensive experiments on varying jailbreak benchmarks across a wide range of LLMs show that IA could consistently and significantly reduce the harmfulness in responses (averagely -48.2{\%} attack success rate). Encouragingly, with our IA, Vicuna-7B even outperforms GPT-3.5 regarding attack success rate. We empirically demonstrate that, to some extent, IA is robust to errors in generated intentions. Further analyses reveal the underlying principle of IA: suppressing LLM`s tendency to follow jailbreak prompts, thereby enhancing safety. |
leng-xiong-2025-towards | 2,025 | Towards Understanding Multi-Task Learning (Generalization) of LLMs via Detecting and Exploring Task-Specific Neurons | While large language models (LLMs) have demonstrated superior multi-task capabilities, understanding the learning mechanisms behind this is still a challenging problem. In this paper, we attempt to understand such mechanisms from the perspective of neurons. Specifically, we detect task-sensitive neurons in LLMs via gradient attribution on task-specific data. Through extensive deactivation and fine-tuning experiments, we demonstrate that the detected neurons are highly correlated with the given task, which we term as task-specific neurons. With these identified task-specific neurons, we delve into two common problems in multi-task learning and continuous learning: Generalization and Catastrophic Forgetting. We find that the overlap of task-specific neurons is strongly associated with generalization and specialization across tasks. Interestingly, at certain layers of LLMs, there is a high similarity in the parameters of different task-specific neurons, and such similarity is highly correlated with the generalization performance. Inspired by these findings, we propose a neuron-level continuous fine-tuning method that only fine-tunes the current task-specific neurons during continuous learning, and extensive experiments demonstrate the effectiveness of the proposed method. Our study provides insights into the interpretability of LLMs in multi-task learning. |
ren-etal-2025-large | 2,025 | Do Large Language Models Mirror Cognitive Language Processing? | Large Language Models (LLMs) have demonstrated remarkable abilities in text comprehension and logical reasoning, indicating that the text representations learned by LLMs can facilitate their language processing capabilities. In neuroscience, brain cognitive processing signals are typically utilized to study human language processing. Therefore, it is natural to ask how well the text embeddings from LLMs align with the brain cognitive processing signals, and how training strategies affect the LLM-brain alignment? In this paper, we employ Representational Similarity Analysis (RSA) to measure the alignment between 23 mainstream LLMs and fMRI signals of the brain to evaluate how effectively LLMs simulate cognitive language processing. We empirically investigate the impact of various factors (e.g., pre-training data size, model scaling, alignment training, and prompts) on such LLM-brain alignment. Experimental results indicate that pre-training data size and model scaling are positively correlated with LLM-brain similarity, and alignment training can significantly improve LLM-brain similarity. Explicit prompts contribute to the consistency of LLMs with brain cognitive language processing, while nonsensical noisy prompts may attenuate such alignment. Additionally, the performance of a wide range of LLM evaluations (e.g., MMLU, Chatbot Arena) is highly correlated with the LLM-brain similarity. |
guan-etal-2025-saged | 2,025 | SAGED: A Holistic Bias-Benchmarking Pipeline for Language Models with Customisable Fairness Calibration | The development of unbiased large language models is widely recognized as crucial, yet existing benchmarks fall short in detecting biases due to limited scope, contamination, and lack of a fairness baseline. SAGED(bias) is the first holistic benchmarking pipeline to address these problems. The pipeline encompasses five core stages: scraping materials, assembling benchmarks, generating responses, extracting numeric features, and diagnosing with disparity metrics. SAGED includes metrics for max disparity, such as impact ratio, and bias concentration, such as Max Z-scores. Noticing that metric tool bias and contextual bias in prompts can distort evaluation, SAGED implements counterfactual branching and baseline calibration for mitigation. For demonstration, we use SAGED on G20 Countries with popular 8b-level models including Gemma2, Llama3.1, Mistral, and Qwen2. With sentiment analysis, we find that while Mistral and Qwen2 show lower max disparity and higher bias concentration than Gemma2 and Llama3.1, all models are notably biased against countries like Russia and (except for Qwen2) China. With further experiments to have models role-playing U.S. presidents, we see bias amplifies and shifts in heterogeneous directions. Moreover, we see Qwen2 and Mistral not engage in role-playing, while Llama3.1 and Gemma2 role-play Trump notably more intensively than Biden and Harris, indicating role-playing performance bias in these models. |
chen-etal-2025-learning | 2,025 | Learning to Reason via Self-Iterative Process Feedback for Small Language Models | Small language models (SLMs) are more efficient, cost-effective, and customizable than large language models (LLMs), though they often underperform in specific areas like reasoning. Past methods for enhancing SLMs' reasoning, such as supervised fine-tuning and distillation, often depend on costly external signals, resulting in SLMs being overly confident with limited supervision signals, thus limiting their abilities. Therefore, this study enables SLMs to learn to reason from self-iterative feedback. By combining odds ratio preference optimization (ORPO), we fine-tune and align SLMs using positive and negative signals generated by themselves. Additionally, we introduce process supervision for rewards in preference alignment by sampling-based inference simulation and process reward models. Compared to Supervised Fine-Tuning (SFT), our method improves the performance of Gemma-2B by 12.43 (Acc) on GSM8K and 3.95 (Pass@1) on MBPP. Furthermore, the proposed method also demonstrated superior out-of-domain generalization capabilities on MMLU{\_}Math and HumanEval. |
cao-etal-2025-rethinking | 2,025 | Rethinking-based Code Summarization with Chain of Comments | Automatic code summarization aims to generate concise natural language descriptions (summary) for source code, which can free software developers from the heavy burden of manual commenting and software maintenance. Existing methods focus on learning a direct mapping from pure code to summaries, overlooking the significant heterogeneity gap between code and summary. Moreover, existing methods lack a human-like re-check process to evaluate whether the generated summaries match well with the code. To address these two limitations, we introduce RBCoSum, a novel framework that incorporates the generated Chain Of Comments (COC) as auxiliary intermediate information for the model to bridge the gap between code and summaries. Also, we propose a rethinking process where a learned ranker trained on our constructed ranking dataset scores the extent of matching between the generated summary and the code, selecting the highest-scoring summary to achieve a re-check process. We conduct extensive experiments to evaluate our approach and compare it with other automatic code summarization models as well as multiple code Large Language Models (LLMs). The experimental results show that RBCoSum is effective and outperforms baselines by a large margin. The human evaluation also proves the summaries generated with RBCoSum are more natural, informative, useful, and truthful. |
feng-he-2025-rgr | 2,025 | RGR-KBQA: Generating Logical Forms for Question Answering Using Knowledge-Graph-Enhanced Large Language Model | In the field of natural language processing, Knowledge Base Question Answering (KBQA) is a challenging task that involves accurately retrieving answers from structured knowledge. Existing methods often face issues when generating query statements using LLMs, as the knowledge introduced may be imprecise and the models themselves may exhibit hallucination problems, leading to low accuracy, particularly when dealing with complex questions. To address these challenges, we introduce a novel semantic parsing approach called RGR-KBQA, which adopts a Retrieve-Generate-Retrieve framework. The first retrieval step introduces factual knowledge from a knowledge graph to enhance the semantic understanding capabilities of LLMs, thereby improving generation accuracy of logical form. The second step uses a fine-tuned model to generate the logical form, and the final step involves unsupervised relation and entity retrieval to further enhance generation accuracy. These two retrieval steps help alleviate the hallucination problems inherent in LLMs. Experimental results show that RGR-KBQA demonstrate promising performance on CWQ and WebQSP datasets. |
azeemi-etal-2025-label | 2,025 | To Label or Not to Label: Hybrid Active Learning for Neural Machine Translation | Active learning (AL) techniques reduce labeling costs for training neural machine translation (NMT) models by selecting smaller representative subsets from unlabeled data for annotation. Diversity sampling techniques select heterogeneous instances, while uncertainty sampling methods select instances with the highest model uncertainty. Both approaches have limitations - diversity methods may extract varied but trivial examples, while uncertainty sampling can yield repetitive, uninformative instances. To bridge this gap, we propose Hybrid Uncertainty and Diversity Sampling (HUDS), an AL strategy for domain adaptation in NMT that combines uncertainty and diversity for sentence selection. HUDS computes uncertainty scores for unlabeled sentences and subsequently stratifies them. It then clusters sentence embeddings within each stratum and computes diversity scores by distance to the centroid. A weighted hybrid score that combines uncertainty and diversity is then used to select the top instances for annotation in each AL iteration. Experiments on multi-domain German-English and French-English datasets demonstrate the better performance of HUDS over other strong AL baselines. We analyze the sentence selection with HUDS and show that it prioritizes diverse instances having high model uncertainty for annotation in early AL iterations. |
yan-etal-2025-llm | 2,025 | LLM Sensitivity Evaluation Framework for Clinical Diagnosis | Large language models (LLMs) have demonstrated impressive performance across various domains. However, for clinical diagnosis, higher expectations are required for LLM`s reliability and sensitivity: thinking like physicians and remaining sensitive to key medical information that affects diagnostic reasoning, as subtle variations can lead to different diagnosis results. Yet, existing works focus mainly on investigating the sensitivity of LLMs to irrelevant context and overlook the importance of key information. In this paper, we investigate the sensitivity of LLMs, i.e. GPT-3.5, GPT-4, Gemini, Claude3 and LLaMA2-7b, to key medical information by introducing different perturbation strategies. The evaluation results highlight the limitations of current LLMs in remaining sensitive to key medical information for diagnostic decision-making. The evolution of LLMs must focus on improving their reliability, enhancing their ability to be sensitive to key information, and effectively utilizing this information. These improvements will enhance human trust in LLMs and facilitate their practical application in real-world scenarios. Our code and dataset are available at https://github.com/chenwei23333/DiagnosisQA. |
chen-etal-2025-unveiling | 2,025 | Unveiling Uncertainty: A Deep Dive into Calibration and Performance of Multimodal Large Language Models | Multimodal large language models (MLLMs) combine visual and textual data for tasks like image captioning and visual question answering. Proper uncertainty calibration is crucial but challenging for reliable use in areas like healthcare and autonomous driving. This paper investigates several MLLMs, focusing on their calibration across various scenarios, including before and after visual fine-tuning as well as before and after multimodal training of the base LLMs. We observed miscalibration in their performance, and at the same time, no significant differences in calibration across these scenarios. We also highlight differences in uncertainty between text and the impact of the integration of these two types of information in uncertainty. To better understand MLLMs' miscalibration and their ability to self-assess uncertainty, we developed the IDK (I don`t know) dataset, which is key for evaluating how they handle unknowns. Our findings reveal that MLLMs tend to give answers rather than admit uncertainty, but this self-assessment improves with prompt adjustments. Finally, to calibrate MLLMs and enhance model reliability, we propose techniques such as temperature scaling and iterative prompt optimization. Our results provide insights into improving MLLMs for effective and responsible deployment in multimodal applications. |
wang-etal-2025-unifying | 2,025 | Unifying Dual-Space Embedding for Entity Alignment via Contrastive Learning | Entity alignment (EA) aims to match identical entities across different knowledge graphs (KGs). Graph neural network-based entity alignment methods have achieved promising results in Euclidean space. However, KGs often contain complex local and hierarchical structures, which are hard to represent in a single space. In this paper, we propose a novel method named as UniEA, which unifies dual-space embedding to preserve the intrinsic structure of KGs. Specifically, we simultaneously learn graph structure embeddings in both Euclidean and hyperbolic spaces to maximize the consistency between embeddings in the two spaces. Moreover, we employ contrastive learning to mitigate the misalignment issues caused by similar entities, where embeddings of similar neighboring entities become too close. Extensive experiments on benchmark datasets demonstrate that our method achieves state-of-the-art performance in structure-based EA. Our code is available at https://github.com/wonderCS1213/UniEA. |
fan-etal-2025-aspect | 2,025 | Aspect-Based Sentiment Analysis with Syntax-Opinion-Sentiment Reasoning Chain | Despite the impressive capabilities of large language models (LLMs) in aspect-based sentiment analysis (ABSA), the role of syntactic information remains underexplored in LLMs. Syntactic structures are known to be crucial for capturing aspect-opinion relationships. To explore whether LLMs can effectively leverage syntactic information to improve ABSA performance, we propose a novel multi-step reasoning framework, the Syntax-Opinion-Sentiment Reasoning Chain (Syn-Chain). Syn-Chain sequentially analyzes syntactic dependencies, extracts opinions, and classifies sentiment. We introduce Syn-Chain into LLMs via zero-shot prompting, and results show that Syn-Chain significantly enhances ABSA performance, though smaller LLM exhibit weaker performance. Furthermore, we enhance smaller LLMs via distillation using GPT-3.5-generated Syn-Chain responses, achieving state-of-the-art ABSA performance. Our findings highlight the importance of syntactic information for improving LLMs in ABSA and offer valuable insights for future research. |
shen-etal-2025-reasoning | 2,025 | Reasoning with Trees: Faithful Question Answering over Knowledge Graph | Recent advancements in large language models (LLMs) have shown remarkable progress in reasoning capabilities, yet they still face challenges in complex, multi-step reasoning tasks. This study introduces Reasoning with Trees (RwT), a novel framework that synergistically integrates LLMs with knowledge graphs (KGs) to enhance reasoning performance and interpretability. RwT reformulates knowledge graph question answering (KGQA) as a discrete decision-making problem, leveraging Monte Carlo Tree Search (MCTS) to iteratively refine reasoning paths. This approach mirrors human-like reasoning by dynamically integrating the LLM`s internal knowledge with external KG information. We propose a real-data guided iteration technique to train an evaluation model that assesses action values, improving the efficiency of the MCTS process. Experimental results on two benchmark KGQA datasets demonstrate that RwT significantly outperforms existing state-of-the-art methods, with an average performance improvement of 9.81{\%}. Notably, RwT achieves these improvements without requiring complete retraining of the LLM, offering a more efficient and adaptable approach to enhancing LLM reasoning capabilities. |
li-etal-2025-revisiting | 2,025 | Revisiting Jailbreaking for Large Language Models: A Representation Engineering Perspective | The recent surge in jailbreaking attacks has revealed significant vulnerabilities in Large Language Models (LLMs) when exposed to malicious inputs. While various defense strategies have been proposed to mitigate these threats, there has been limited research into the underlying mechanisms that make LLMs vulnerable to such attacks. In this study, we suggest that the self-safeguarding capability of LLMs is linked to specific activity patterns within their representation space. Although these patterns have little impact on the semantic content of the generated text, they play a crucial role in shaping LLM behavior under jailbreaking attacks. Our findings demonstrate that these patterns can be detected with just a few pairs of contrastive queries. Extensive experimentation shows that the robustness of LLMs against jailbreaking can be manipulated by weakening or strengthening these patterns. Further visual analysis provides additional evidence for our conclusions, providing new insights into the jailbreaking phenomenon. These findings highlight the importance of addressing the potential misuse of open-source LLMs within the community. |
schoene-etal-2025-lexicography | 2,025 | Lexicography Saves Lives (LSL): Automatically Translating Suicide-Related Language | Recent years have seen a marked increase in research that aims to identify or predict risk, intention or ideation of suicide. The majority of new tasks, datasets, language models and other resources focus on English and on suicide in the context of Western culture. However, suicide is global issue and reducing suicide rate by 2030 is one of the key goals of the UN`s Sustainable Development Goals. Previous work has used English dictionaries related to suicide to translate into different target languages due to lack of other available resources. Naturally, this leads to a variety of ethical tensions (e.g.: linguistic misrepresentation), where discourse around suicide is not present in a particular culture or country. In this work, we introduce the {\textquoteleft}Lexicography Saves Lives Project' to address this issue and make three distinct contributions. First, we outline ethical consideration and provide overview guidelines to mitigate harm in developing suicide-related resources. Next, we translate an existing dictionary related to suicidal ideation into 200 different languages and conduct human evaluations on a subset of translated dictionaries. Finally, we introduce a public website to make our resources available and enable community participation. |
Subsets and Splits