Language models encounter unique difficulties when analyzing interdisciplinary scientific literature. In biophysics, a faithful answer must ground observed data in source evidence, interpret it with a quantitative physics model, and connect it to a biological mechanism. To meet this need we introduce BioPhys-Bridge, a benchmark dataset focused on evidence‑grounded scientific reasoning. Each case supplies evidence blocks, stable evidence IDs, numeric values, units, equations, assumptions, mechanisms, and subsequent decisions as grounding targets for question answering (QA) and retrieval‑augmented generation (RAG). The initial release comprises 500 cases, 1,517 agent‑facing tasks, covering six biological domains and nine families of physical models, with three sparse families reserved for future expansion. All cases pass strict quality gates on schema, evidence integrity, quantitative grounding, source license, deduplication, and unit normalization, and 81 cases have been reviewed and annotated by domain experts. Preliminary results show DeepSeek‑V4‑Flash achieving the highest evidence‑ID $F_1$ score of 0.360, followed by Qwen3.7‑Max (0.316) and GPT‑4o‑mini (0.294). BioPhys-Bridge serves as an interdisciplinary benchmark for assessing attribution, faithfulness, hallucination reduction, and complex multi‑step biological experiment design reasoning. Future work will enlarge the dataset’s size and complexity and conduct comprehensive evaluations. Code and data are publicly available on GitHub and Hugging Face.
Review: This benchmark, with its fine‑grained evidence annotations and quantitative details, offers a valuable platform for measuring the reliability of large models in cross‑disciplinary scientific reasoning.