BioPhys-Bridge: A Benchmark for Scientific Reasoning in Physics-Grounded Biological Research

Summary

Language models must connect source evidence, quantitative physics, and biological mechanisms when analyzing biophysics literature. BioPhys-Bridge is a new benchmark for evidence-grounded question answering and retrieval-augmented generation, with each case specifying evidence blocks, stable evidence IDs, quantitative values, units, equations, assumptions, mechanisms, and possible next decisions. Its initial release contains 500 cases and 1,517 agent-facing tasks spanning six biological domains and nine physical model families, with three sparse families reserved for future expansion. The dataset applies schema, evidence-integrity, quantitative-grounding, source-license, duplicate, and unit-normalization checks, alongside domain-expert review and annotation for 81 cases. In preliminary evaluation, DeepSeek-V4-Flash achieved the highest evidence-ID F1 score at 0.360, followed by Qwen3.7-Max at 0.316 and GPT-4o-mini at 0.294. The benchmark is designed to assess attribution, faithfulness, hallucination reduction, and biological experiment design under multi-step scientific reasoning. The authors say the code and data are available through GitHub and Hugging Face, and plan to expand the dataset and conduct broader evaluations.