New benchmark evaluates language models on interdisciplinary biophysics reasoning

Researchers have introduced BioPhys-Bridge, a benchmark dataset designed to evaluate language models on interdisciplinary scientific reasoning in biophysics. The dataset focuses on tasks that require grounding answers in source evidence, interpreting data through quantitative physics models, and linking findings to biological mechanisms. Each case in the benchmark includes structured evidence blocks, stable evidence IDs, quantitative values with units, equations, assumptions, mechanisms, and next-step decisions to guide question answering and retrieval-augmented generation.

The initial release of BioPhys-Bridge comprises 500 cases and 1,517 agent-facing tasks, spanning six biological domains and nine physical model families. Three additional sparse model families are reserved for future expansion. The dataset enforces strict quality controls, including schema validation, evidence integrity checks, quantitative grounding, source-license verification, duplicate removal, and unit normalization. Domain experts reviewed and annotated 81 cases to ensure reliability.

Preliminary evaluations on BioPhys-Bridge show DeepSeek-V4-Flash achieving the highest evidence-ID F1 score of 0.360, followed by Qwen3.7-Max at 0.316 and GPT-4o-mini at 0.294. These results highlight the benchmark’s utility in measuring model performance on evidence attribution and faithfulness in complex scientific reasoning tasks.

The benchmark is positioned as a tool for assessing attribution, reducing hallucinations, and improving biological experiment design in language models. Future work includes expanding the dataset’s size and complexity and conducting comprehensive evaluations across additional models and domains.

Code and data for BioPhys-Bridge are publicly available on GitHub and Hugging Face, enabling researchers to reproduce results and extend the benchmark with new cases or tasks.