Principia: Test-Time Hypotheses to Principle Construction for Scientific Discovery
In brief
Principia is an LLM-based framework that constructs comparative causal hypotheses at test time from related inputs and observed outcomes, then generalizes recurring relationships into scientific principles that are continually revised as new observations arrive. Across benchmarks in molecular optimization, protein engineering, gene perturbation, and molecular property prediction, it outperforms baselines relying on internal knowledge, external sources, or experience summaries, and its generated principles are plausible and consistent with the scientific literature. Code: https://github.com/DongkiKim95/Principia.
Why it matters
Principia turns related inputs and observations into comparative causal hypotheses, generalizes them into updatable scientific principles, and beats the listed baselines on four biomedical benchmark types.
