BioPhys-Bridge
An interdisciplinary scientific reasoning benchmark for biophysics literature. It evaluates models on attribution, faithfulness, and hallucination suppression in multi-step reasoning using evidence-based tasks. The benchmark supports both question answering (QA) and retrieval-augmented generation (RAG) task formats.
Timeline (last 90 days)
BioPhys-Bridge is an evidence-based scientific reasoning evaluation benchmark for biophysics literature.
BioPhys-Bridge breaks down each case into a series of structured grounding targets, including evidence blocks, evidence IDs, quantitative values, units, formulas, prerequisite assumptions, biological mechanisms, and next-step decision directions.
BioPhys-Bridge can simultaneously support both question answering (QA) and retrieval-augmented generation (RAG) tasks.
The first release version of BioPhys-Bridge contains 500 cases and 1,517 agent-oriented tasks, covering six major biological domains and nine physical model families.
Three of the nine physical model families belong to sparse categories and have been reserved for future expansion.
BioPhys-Bridge's RAG tasks require models to precisely locate the specific passages or values from an evidence block library that support the answer, and an incorrect evidence ID is judged as a failure even if the answer is correct.
BioPhys-Bridge sets multiple quality thresholds, including Schema specification, evidence completeness, quantitative grounding, source licensing, deduplication, and unit normalization.
The research team invited domain experts to manually review and annotate 81 cases within BioPhys-Bridge.
In the BioPhys-Bridge evaluation, DeepSeek-V4-Flash achieved the highest score with an evidence ID F1 score of 0.360.
In the BioPhys-Bridge evaluation, Qwen3.7-Max's evidence ID F1 score is 0.316.
3 more timeline events
All Facts (13)
Even the best-performing model achieves an evidence ID F1 of only around 0.36, reflecting that current large models still have significant room for improvement on cross-disciplinary reasoning tasks requiring precise evidence localization.
70%UnverifiedBioPhys-Bridge breaks down each case into a series of structured grounding targets, including evidence blocks, evidence IDs, quantitative values, units, formulas, prerequisite assumptions, biological mechanisms, and next-step decision directions.
50%UnverifiedBioPhys-Bridge can simultaneously support both question answering (QA) and retrieval-augmented generation (RAG) tasks.
50%UnverifiedThree of the nine physical model families belong to sparse categories and have been reserved for future expansion.
50%UnverifiedBioPhys-Bridge's RAG tasks require models to precisely locate the specific passages or values from an evidence block library that support the answer, and an incorrect evidence ID is judged as a failure even if the answer is correct.
50%UnverifiedBioPhys-Bridge sets multiple quality thresholds, including Schema specification, evidence completeness, quantitative grounding, source licensing, deduplication, and unit normalization.
50%UnverifiedThe research team invited domain experts to manually review and annotate 81 cases within BioPhys-Bridge.
50%UnverifiedIn the BioPhys-Bridge evaluation, DeepSeek-V4-Flash achieved the highest score with an evidence ID F1 score of 0.360.
50%UnverifiedIn the BioPhys-Bridge evaluation, Qwen3.7-Max's evidence ID F1 score is 0.316.
50%UnverifiedBioPhys-Bridge is an evidence-based scientific reasoning evaluation benchmark for biophysics literature.
50%UnverifiedBioPhys-Bridge emphasizes the evaluation of capabilities including attribution, faithfulness, hallucination suppression, and biological experiment design.
50%UnverifiedThe code and data related to BioPhys-Bridge have been made publicly available on a GitHub repository and Hugging Face.
50%UnverifiedThe first release version of BioPhys-Bridge contains 500 cases and 1,517 agent-oriented tasks, covering six major biological domains and nine physical model families.
50%