--- title: Inkling emoji: 🌐 colorFrom: indigo colorTo: yellow # python_version: 3.10 sdk: gradio sdk_version: 5.29.0 app_file: app.py pinned: true license: agpl-3.0 short_description: Use AI to find obvious research links in unexpected places. datasets: - nomadicsynth/arxiv-dataset-abstract-embeddings models: - nomadicsynth/research-compass-arxiv-abstracts-embedding-model --- # Inkling: Bridging the Unconnected in Scientific Literature ![Inkling Logo - A cartoon squid with a smile. The human-like brain is visible.](https://huggingface.co/spaces/nomadicsynth/inkling/resolve/main/inkling-logo.png) **Inkling** is an experimental bridge-finding engine for scientific literature, built to uncover *latent connections* between research papers—relationships that are obvious in hindsight but buried under the sheer volume of modern research. It’s inspired by the work of **Don R. Swanson**, the visionary who discovered the link between *fish oil* and *Raynaud’s syndrome* using nothing but manual literature analysis. Today, we call this approach **Literature-Based Discovery** - and Inkling is our attempt to automate it with modern NLP. --- ## The Problem: Lost in the Literature The scientific literature is growing exponentially, but human researchers can only read so much. As Sabine Hossenfelder explained in her 2024 YouTube video ["AIs Predict Research Results Without Doing Research"](https://www.youtube.com/watch?v=Qgrl3JSWWDE), even experts miss critical connections because no one has time to read everything. Swanson’s 1986 discovery of the fish oil–Raynaud’s link was a wake-up call: the knowledge existed in plain sight, but the papers were siloed. Inkling is our attempt to fix that. --- ## The Vision: A Bridge-Finding Machine Inkling isn’t just a search engine. It’s a **hypothesis generator**. It learns to recognize *intermediate concepts* that connect seemingly unrelated papers—like Swanson’s "blood viscosity" bridge. The model is built to: - **Find indirect links** between papers that don’t cite each other. - **Surface connections** that feel obvious once explained but are buried in the noise. - **Scale** to the entire arXiv corpus and beyond. --- ## How It Works ### Model Architecture - **Base Model**: A `SentenceTransformer` using **Llama-7B** as its base (with frozen weights) and a dense embedding head. - **Training**: - v1: Trained on a synthetic dataset of randomly paired papers, rated for conceptual overlap. - v2 (in progress): Focused on *bridge detection*, using prompts to explicitly identify intermediate concepts (e.g., "What connects these two papers?"). - **Embedding Strategy**: - Dense vector representations of abstracts. - FAISS for fast approximate nearest-neighbor search. ### Dataset Philosophy - v1: Random paper pairs rated for generic "relevance" (too broad, limited bridge detection). - v2: Focus on **explicit bridge extraction** using LLM-generated triplets (e.g., "Paper A → Bridge Concept → Paper B"). --- ## The Inspiration This project was born from a **nerd-sniping moment** after watching Sabine Hossenfelder’s video on AI’s ability to predict neuroscience results without experiments. That led to three key influences: ### 1. **Swanson’s "Undiscovered Public Knowledge"** Swanson’s 1986 paper showed that the fish oil–Raynaud’s link existed in the literature for decades—it just took a human to connect the dots. Inkling automates this process. ### 2. **Tshitoyan et al. (2019): Word Embeddings in Materials Science** Their work demonstrated that unsupervised embeddings could predict future material discoveries from latent knowledge. Inkling applies this idea to *conceptual bridges* in all scientific fields. ### 3. **Luo et al. (2024): LLMs Beat Human Experts** This study showed that a 7B LLM (like Mistral) could outperform neuroscientists in predicting experimental outcomes. Inkling leverages this power to find connections even domain experts might miss. --- ## What It Can Do (and What’s Next) ### Current Capabilities - Embed arXiv abstracts into dense vectors. - Search for papers with conceptual overlap (50% relevance in top-10/25 queries, per manual testing). - Visualize results in a Gradio interface with FAISS-powered speed. ### Roadmap - **v2**: Train on LLM-generated bridge triplets (e.g., "Paper A → Blood Viscosity → Paper B"). - **Gradio Enhancements**: - Interactive bridge visualization (D3.js or Plotly). - User feedback loop for improving the model. - **Automated Updates**: Embed new arXiv papers nightly. - **Domain-Specific Tools**: - Drug repurposing mode (e.g., "Find new uses for aspirin"). - Interdisciplinary connection finder (e.g., "How does physics inform AI research?"). --- ## Why This Matters Inkling is **not** a polished product—it’s a chaotic, ADHD-fueled experiment in democratizing scientific discovery. It’s for: - Researchers drowning in paper overload. - Interdisciplinary thinkers who thrive on unexpected connections. - Anyone who’s ever thought, *"I could’ve thought of that!"* after a breakthrough. As Sabine Hossenfelder put it: *"The future of research isn’t in doing more experiments—it’s in connecting the dots we already have."* - Citation needed. --- ## Status - **Model**: v1 (proof of concept, 50-50 if it does anything or my brain is just playing tricks). - **Dataset**: v1 (random pairs, too broad). v2 (in planning, focused on bridge detection). - **Interface**: Gradio-powered demo with FAISS backend. - **Next Steps**: Refine training data, automate updates, and scale to all of arXiv. --- ## Credits - **Inspiration**: Sabine Hossenfelder’s ["AIs Predict Research Results" video](https://www.youtube.com/watch?v=Qgrl3JSWWDE). - **Foundational Work**: Don R. Swanson, V. Tshitoyan, X. Luo. - **Model Architecture**: Llama-7B + SentenceTransformer. --- ## Try It [**Live Demo**](https://nomadicsynth-research-compass.hf.space) *Paste an abstract, find a bridge, and see if the connection feels obvious in hindsight.* 🚀 --- **This is a work in progress. Feedback, ideas, and nerd-sniped collaborators are welcome.**