BioPhys-Bridge Benchmark Tests Interdisciplinary Scientific Reasoning: DeepSeek-V4-Flash Leads with F1 = 0.360

It is an evaluation dataset for evidence-grounded scientific reasoning over biophysics literature, containing evidence blocks, stable evidence IDs, quantitative values, units, equations, assumptions, and mechanisms, with 500 cases and 1,517 agent-facing tasks.

DeepSeek-V4-Flash achieved the highest evidence-ID F1 score at 0.360, versus 0.316 for Qwen3.7-Max and 0.294 for GPT-4o-mini, indicating a relative advantage in linking answers to source evidence.

Because the top score of 0.360 means current models fail in more than 60% of precise attribution tasks, a direct indicator of hallucination risk in sensitive scientific contexts.

Yes, especially teams in biotech, health, energy, and environment, since it offers a free, reproducible tool to evaluate RAG models before deploying them in scientific research workflows.