Reproduce “Medical Interpretability and Knowledge Maps of Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Razvan Marinescu, Victoria-Elisabeth Gruber, Diego Fajardo
Researchers at Lumos AI systematically investigated how Large Language Models represent medical knowledge by applying four interpretability techniques across five open-source models and several medical knowledge domains. The study developed "LLM knowledge maps" that pinpoint specific layers responsible for different medical concepts, revealing non-linear and discontinuous age encoding, circular patterns in disease progression, and drug clustering based on medical specialty.
Key insights
- Medical knowledge representations in LLMs are often complex and non-linear, with age encoding being gender-dependent and discontinuous at certain thresholds (e.g., 17-18 years old).
- LLMs can form intricate, non-monotonic semantic structures, such as circular patterns for dynamic concepts like disease progression, where late stages can be closer to early stages in latent space.
- The organization of medical knowledge within LLMs can reflect pragmatic clinical contexts, as evidenced by drugs clustering more effectively by medical specialty than by their mechanism of action.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.