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

Qingyang Xu

cs.AISep 15, 2026

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

Authors: Qingyang Xu

Organizations: Independent Researcher Shanghai, China

Abstract

Language models face unique challenges in analyzing interdisciplinary scientific research literature. In biophysics research, faithful answers require grounding observed data in source evidence, interpreting it through a quantitative physics model, and linking it to a biological mechanism. To address this challenge, we introduce BioPhys-Bridge, a novel benchmark dataset for evidence-grounded scientific reasoning over biophysical literature. Each case contains evidence blocks, stable evidence IDs, quantitative values, units, equations, assumptions, mechanisms, and next decisions as grounding targets for question answering (QA) and retrieval-augmented generation (RAG). The initial release contains 500 cases, 1,517 agent-facing tasks, and covers six biological domains and nine physical model families, including three sparse families reserved for future expansion. We enforce strict quality gates for all cases in schema, evidence-integrity, quantitative-grounding, source-license, duplicate, unit-normalization, with domain expert review and annotation for 81 cases. Preliminary evaluations show that DeepSeek-V4-Flash obtain the highest evidence-ID $F_1$ score (0.360), followed by Qwen3.7-Max (0.316) and GPT-4o-mini (0.294). BioPhys-Bridge is an interdisciplinary benchmark for evaluating attribution, faithfulness, hallucination reduction, and biological experiment design with complex, multi-step scientific reasoning. Future works will increase the size and complexity of the dataset and perform comprehensive evaluations. Code and data are available in the GitHub repository and on Hugging Face.

Figures & tables

Benchmark family	Core unit	Primary strength	Gap addressed by BioPhys-Bridge
PaperQA, LitQA ( Lala et al., 2023 ) ; LAB-Bench ( Laurent et al., 2024 )	Scientific papers and research questions	Literature-grounded QA, biology research skills, RAG evaluation	Equations, units, assumptions, and mechanism links are not first-class case fields.
ScienceQA ( Lu et al., 2022 ) ; SciBench ( Wang et al., 2024b ) ; GPQA ( Rein et al., 2024 )	Science questions or problems	Scientific knowledge, explanations, expert difficulty	No explicit evidence-to-equation-to-mechanism case object.
MathVista ( Lu et al., 2024 ) ; MMMU ( Yue et al., 2024 ) ; SPIQA ( Pramanick et al., 2024 )	Multimodal questions and paper figures	Visual reasoning and scientific document QA	Physical models and biological mechanisms are not first-class grounding targets.
PDEBench ( Takamoto et al., 2022 ) ; BubbleML ( Hassan et al., 2023 )	Simulation data and physical benchmark tasks	Physics-grounded prediction with controlled ground truth	Focuses on numerical modeling rather than literature-grounded mechanism decisions.
BioPhys-Bridge (ours)	Evidence-linked scientific reasoning cases	Evidence IDs, quantitative values, equations, assumptions, mechanisms, and next decisions	Designed for evidence-grounded equation-to-mechanism reasoning over biophysical literature.

Table 1: Positioning BioPhys-Bridge against recent benchmark families. The distinguishing unit is not a question alone, but a structured case linking source evidence, physical model, quantitative evidence, biological mechanism, and scientific decision.

Figure 1: Each case as an evidence-linked bridge from source-paper evidence to quantitative values, physical-model assumptions, mechanistic interpretation, and an agent-facing scientific decision task.

Field	Purpose
case_id	Stable identifier for an evidence-grounded case.
domain , bridge_type	Biological setting and domain-specific physics-to-mechanism bridge.
source	DOI/PMCID, title, license, source URL, and MinerU parse identifier.
evidence[]	Text, table, formula, and figure/caption blocks with stable evidence IDs.
quantitative_evidence[]	Metric/value/unit records citing evidence IDs.
biophysical_model	Model name, family, equation, variables, assumptions, validity conditions.

Table 2: Dataset schema of BioPhys-Bridge. Evidence IDs, quantitative values, equations, assumptions, mechanisms, and decisions are explicit grounding targets.

Figure 2: Curation pipeline. Open-access papers are parsed with MinerU, normalized into evidence blocks, structured into case records, and filtered by validation and quality gates before release.

Release item	Value
Cases	500
Train / validation / test	400 / 50 / 50
Agent tasks	1,517
Failure/revision cases	107
Expert-annotation cases	81
Mean modalities per case	2.944

Table 3: BioPhys-Bridge v1 release at a glance.

Biological domain	Cases
protein_ligand_binding	188
systems_biology_dynamics	126
conformational_dynamics_allostery	91
biomolecular_phase_separation	52
enzyme_kinetics	22
protein_stability_thermodynamics	21

Table 4: Biological-domain coverage in the 500-case release.

Task type	Count
mechanism_from_evidence	480
next_experiment_design	403
derivation	355
discrepancy_explanation	279

Table 5: Agent-task distribution.

Gate	Check	Release result
Schema validation	Pydantic and generated JSON Schema	500 / 500 valid
Evidence integrity	Cited IDs resolve to evidence blocks	100% coverage
Quantitative grounding	Values appear in cited evidence text	100% coverage
Source compliance	Release-compatible source licenses	490 CC-BY-4.0, 10 CC0-1.0
Unit normalization	Raw and normalized units retained	100% success
Duplicate check	Stable case IDs are unique	0 duplicates

Table 6: Release quality gates. The deterministic audit notes are a supporting traceability aid; only about 2% of cases have an implemented closed-form relation-level check, so it is not used as a global correctness metric.

Model	Provider	Evidence-ID F1	Answer token F1	Empty EID outputs
Lexical retrieval	deterministic	0.188 [0.151, 0.226]	0.058 [0.053, 0.065]	0/154
DeepSeek v4 Flash	DeepSeek	0.360 [0.309, 0.412]	0.193 [0.173, 0.214]	0/154
Qwen 3.7 Max	OpenRouter	0.316 [0.273, 0.359]	0.129 [0.115, 0.143]	1/154
GPT-4o-mini	OpenRouter	0.294 [0.251, 0.339]	0.163 [0.149, 0.177]	0/154
Claude Opus 4.8	OpenRouter	0.238 [0.205, 0.272]	0.095 [0.083, 0.107]	18/154
Claude Sonnet 5	OpenRouter	0.224 [0.190, 0.258]	0.077 [0.068, 0.086]	21/154

Table 7: Preliminary de-leaked no-scaffold evaluation on the identical 154-task held-out test set. Values in brackets are bootstrap 95% confidence intervals over tasks. Empty EID outputs counts predictions with no parsed supporting_evidence_ids ; this is an output-compliance measure as well as an attribution failure mode.

Task type	N	Lex.	Flash	Qwen	GPT-mini
derivation	39	0.192	0.475	0.380	0.346
discrepancy	31	0.201	0.404	0.379	0.350
mechanism	46	0.188	0.352	0.311	0.306
next experiment	38	0.173	0.216	0.203	0.181

Table 8: Held-out evidence-ID F1 by task type for the lexical floor and the three strongest LLM rows in Table 7 . Each task-type cell contains only 31–46 tasks, so these point estimates should be read as noisier diagnostics rather than stable per-type rankings.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: A real BioPhys-Bridge case illustrates the benchmark object: source provenance, evidence-linked quantitative measurements, a physical model, mechanism interpretation, caveats, and an agent-facing task are stored together rather than flattened into a standalone question.

Archived release stage	Count
Reviewed source records	500
Schema-valid source records	500
Exported case records	500
Deterministic train split	400
Deterministic validation split	50
Deterministic test split	50

Appendix

Table 9: Release audit funnel recorded in the public metadata. Future dataset versions should additionally archive exploratory source-pool and parser-failure counts to support a full rejection funnel.

Explore similar work

BioAlchemy: Distilling Biological Literature into Reasoning-Ready Reinforcement Learning Training Data
Apr 3, 2026Brian Hsu, Ozan Gökdemir, Carlo Siebenschuh +7Scientific Reasoning in Language Models Scientific QA
Department of Computer Science, University of Chicago, Chicago, IL 60637 · Data Science and Learning Division, Argonne National Laboratory, Lemont, IL 60439
InfiniteScienceGym: An Unbounded, Procedurally-Generated Benchmark for Scientific Analysis
Apr 14, 2026Oliver Bentham, Vivek SrikumarLLM Evaluation Scientific QA
Kahlert School of Computing University of Utah
BRIDGE: Benchmark for multi-hop Reasoning In long multimodal Documents with Grounded Evidence
Mar 9, 2026Biao Xiang, Soyeon Caren Han, Yihao DingMulti-Hop QA Multimodal QA
The University of Melbourne Melbourne, Australia · The University of Western Australia Perth, Australia

Apr 3, 2026J/K move · Enter open · S save

BioAlchemy: Distilling Biological Literature into Reasoning-Ready Reinforcement Learning Training Data

Brian Hsu, Ozan Gökdemir, Carlo Siebenschuh, Bruce Parrello+6

Department of Computer Science, University of Chicago, Chicago, IL 60637 · Data Science and Learning Division, Argonne National Laboratory, Lemont, IL 60439

Scientific Reasoning in Language Models Scientific QA RL for Language Model Reasoning

Open paper AskPDF arXiv

Despite the large corpus of biology training text, the impact of reasoning models on biological research generally lags behind math and coding. In this work, we show that biology questions from current large-scale reasoning datasets do not align well with modern research topic distributions in biology, and that this topic imbalance may negatively affect performance. In addition, we find that methods for extracting challenging and verifiable research problems from biology research text are a critical yet underdeveloped ingredient in applying reinforcement learning for better performance on biology research tasks. We introduce BioAlchemy, a pipeline for sourcing a diverse set of verifiable question-and-answer pairs from a scientific corpus of biology research text. We curate BioAlchemy-345K, a training dataset containing over 345K scientific reasoning problems in biology. Then, we demonstrate how aligning our dataset to the topic distribution of modern scientific biology can be used with reinforcement learning to improve reasoning performance. Finally, we present BioAlchemist-8B, which improves over its base reasoning model by 9.12% on biology benchmarks. These results demonstrate the efficacy of our approach for developing stronger scientific reasoning capabilities in biology. The BioAlchemist-8B model is available at: https://huggingface.co/BioAlchemy.

Abstract

Explore similar work

BioAlchemy: Distilling Biological Literature into Reasoning-Ready Reinforcement Learning Training Data

InfiniteScienceGym: An Unbounded, Procedurally-Generated Benchmark for Scientific Analysis

BRIDGE: Benchmark for multi-hop Reasoning In long multimodal Documents with Grounded Evidence