TrustMed-RL: Long-Horizon Reinforcement Learning for Evidence-Grounded Clinical Diagnosis

Wenxin Zhan; Yizheng Jiao; Haifeng Song; Shuai Xu; Chencheng Pan; Jiayi Feng; Anjie Xie

cs.AIOct 3, 2026

TrustMed-RL: Long-Horizon Reinforcement Learning for Evidence-Grounded Clinical Diagnosis

Authors: Wenxin Zhan, Yizheng Jiao, Haifeng Song, Shuai Xu, Chencheng Pan, Jiayi Feng, Anjie Xie

Organizations: Department of Information Systems and Statistics, Baruch College · Department of Computer Science, University of North Carolina at Chapel Hill · Tsinghua University · New York University · Zhongshan Hospital (Xiamen), Fudan University

Abstract

Medical language models can produce correct diagnoses despite incomplete investigations and unsupported reasoning. To support long-horizon, evidence-grounded diagnosis, we introduce \textbf{TrustMed-RL}. Built from PubMed rare-disease cases and over 24,000 manually annotated image panels, it integrates interviews, examinations, testing, specialist consultation, and literature search through state-dependent actions. Our 8B vision--language policy, trained with clinically adapted GiGPO and coverage-adjusted diagnostic rewards, achieves 37.1% diagnostic accuracy on 2,500 evaluation cases, outperforming all evaluated open-weight baselines and improving over supervised fine-tuning by 12.4 percentage points.. When success additionally requires acquiring at least 50% of supporting test evidence, TrustMed-RL achieves 32.5%, exceeding GPT-4o by 6.8 percentage points. Furthermore, it surpasses all evaluated baselines on MTMedDialog and multiple larger 27--32B models on AgentClinic. In physician review of 200 diagnostically accepted test-set trajectories, 83.0% receive evidential-grounding scores of 4--5 out of 5. Physicians' assessments suggest that these diagnostic trajectories are trustworthy and aligned with human diagnostic reasoning.

Figures & tables

Proprietary (closed-source) models
Model	Acc	Acc strict	Prec.	Rec.	Acc@50	Strict@50	$\Delta_{\mathrm{Acc}}$ (%)	$\Delta_{\mathrm{strict}}$ (%)
GPT-5	$54.2$	$22.0$	$42.9$	$50.0$	$43.2$	$17.7$	$20.3$	$19.5$
GPT-4o	$37.9$	$20.8$	$34.9$	$36.0$	$25.7$	$14.3$	$32.3$	$31.3$
Open-weight models
TrustMed-RL (GiGPO, 8B)	$\mathbf{37.1}_{\scriptscriptstyle\pm 1.1}$	$15.3_{\scriptscriptstyle\pm 1.0}$	$\mathbf{31.8}_{\scriptscriptstyle\pm 1.1}$	$\mathbf{34.6}_{\scriptscriptstyle\pm 1.1}$	$\mathbf{32.5}_{\scriptscriptstyle\pm 1.3}$	$\mathbf{13.5}_{\scriptscriptstyle\pm 0.6}$	$12.3$	$11.9$
TrustMed-RL (GRPO, 8B)	$34.5_{\scriptscriptstyle\pm 1.6}$	$13.6_{\scriptscriptstyle\pm 1.9}$	$28.9_{\scriptscriptstyle\pm 1.9}$	$32.1_{\scriptscriptstyle\pm 1.5}$	$30.8_{\scriptscriptstyle\pm 1.3}$	$12.2_{\scriptscriptstyle\pm 1.6}$	$\mathbf{10.7}$	$\mathbf{10.9}$

Table 1: Diagnostic performance on 2,500 clean cases. Scores are percentages;

\pm

denotes reported standard deviations. Acc and Strict require recovery of every reference condition under core-or-full and full matching, respectively. Acc@50 and Strict@50 additionally require at least 50% test-findings coverage. Bold marks the best open-weight mean in each column. Complete comparisons are provided in Appendix Table 5 .

Proprietary (closed-source) models
Model	Order	Findings	Prod.	Workup	Search	Invalid	Modality Catch	Anatomy Catch	False Alarm
GPT-5	$72.4$	$62.8$	$8.5$	$12.8$	$63.2$	$0.0$	$97.8$	$88.7$	$4.8$
GPT-4o	$62.0$	$53.5$	$6.9$	$10.9$	$75.6$	$0.0$	$95.4$	$85.7$	$6.4$
Open-weight models
TrustMed-RL (GiGPO, 8B)	$83.9_{\scriptscriptstyle\pm 4.0}$	$73.2_{\scriptscriptstyle\pm 3.9}$	$8.5_{\scriptscriptstyle\pm 0.4}$	$17.8_{\scriptscriptstyle\pm 0.8}$	$93.0_{\scriptscriptstyle\pm 2.5}$	$4.0_{\scriptscriptstyle\pm 1.1}$	$82.2_{\scriptscriptstyle\pm 1.1}$	$52.6_{\scriptscriptstyle\pm 4.2}$	$2.2_{\scriptscriptstyle\pm 0.8}$
TrustMed-RL (GRPO, 8B)	$\mathbf{84.7}_{\scriptscriptstyle\pm 1.3}$	$\mathbf{73.9}_{\scriptscriptstyle\pm 1.1}$	$\mathbf{9.1}_{\scriptscriptstyle\pm 0.2}$	$18.3_{\scriptscriptstyle\pm 0.6}$	$89.3_{\scriptscriptstyle\pm 4.7}$	$2.6_{\scriptscriptstyle\pm 0.2}$	$85.4_{\scriptscriptstyle\pm 7.9}$	$60.6_{\scriptscriptstyle\pm 2.4}$	$2.4_{\scriptscriptstyle\pm 0.5}$

Table 2: Evidence acquisition and consultation behavior. Order and Findings are mean coverage over 2,452 eligible clean cases. Prod. counts actions yielding new facts, images, or findings; Workup counts workup turns per clean episode. Search is the percentage of episodes with search; Invalid is the percentage of invalid turns. Modality and Anatomy are exposure-conditioned mismatch catch rates. False alarm is measured on matched clean-image gate-replay controls. All metrics except Prod. and Workup are percentages. Complete comparisons appear in Appendix Table 6 .

	Solvable				Unsolvable
Model / variant	ACC	Strict	ACC@3	Strict@3	ACC	Strict	ACC@3	Strict@3	Reference	$\Delta$ solvable ACC [95% CI]
2B Base	2.27	0.60	6.21	1.65	0.44	0.09	1.07	0.27	-	-
2B + SFT	14.57	4.33	31.62	11.20	4.00	0.98	9.33	2.93	2B Base	+12.30 [11.28, 13.33]
8B Base	18.40	7.39	36.71	15.95	4.18	0.71	10.40	2.13	-	-
8B + SFT	27.89	10.40	48.94	21.65	7.64	2.04	17.33	5.07	8B Base	+9.49 [8.03, 10.92]
TrustMed-RL w/o step credit (GRPO)	39.98	17.77	60.00	32.28	11.02	3.91	21.87	8.53	8B SFT	+12.09 [10.59, 13.62]

Table 3: Ablation study. Accuracy values are percentages; gains are percentage points relative to the named reference. ACC@3 and Strict@3 measure the proportion of cases correct in at least one of three evaluations under the corresponding criterion; each case contributes once. Base denotes Qwen3-VL Instruct. All RL variants use SFT initialization. Brackets show paired-case 95% intervals; the four SFT/RL gains also survive Bonferroni adjustment.

Appendix figures & tables47 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Case-conditioned clinical POMDP. The environment maintains the latent state $s_{t}$ and releases a multimodal observation $o_{t}$ to the physician. The physician selects an action $a_{t}$ using its observable interaction context $h_{t}$ , and the environment updates its state through $P$ . The available actions depend on the current interaction phase $\varphi_{t}$ , remaining budgets $b_{t}$ , and evidence and integrity states. The patient simulator and clinical tools are environment components; only the physician policy $\pi_{\theta}$ is optimized. $R(\tau)$ denotes the terminal outcome reward.

Figure 2: Training framework. (a) Teacher-generated SFT. Teacher models interact with the environment to generate multi-turn consultation trajectories. Quality-filtered trajectories form $\mathcal{D}_{\mathrm{SFT}}$ ; the resulting SFT policy initializes $\pi_{\theta}$ and serves as the frozen KL reference $\pi_{\mathrm{ref}}$ . (b) Adaptive GiGPO. A dynamic sampler combines outcome dispersion, staleness, and exploration to select instances admitted by the interaction curriculum. Each instance produces $K$ trajectories under shared initial conditions and budgets. Episode-relative and step-relative advantages jointly guide policy optimization. Observed outcomes update sampling statistics and admission estimates, while a staged schedule increases the interaction-budget ceilings throughout the optimization steps.

Terminal outcome	$R_{i}$
Diagnosis, subject to the overrides below	$R_{i}^{\mathrm{dx}}$
Incorrect diagnosis with no workup action	$-0.05$
Diagnosis after accepting a corrupted image batch	$-0.10$
Evidence Corruption detected with valid axis attribution	$0.90$
Evidence Corruption detected with invalid or missing axis attribution	$0.65$
False or unsupported evidence corruption alarm	$-0.10$

Appendix

Table 4: Terminal rewards used in training. Diagnosis overrides are applied in the order described in the text. Invalid-action costs are separate.

Proprietary (closed-source) models
Model	Acc	Acc strict	Prec.	Rec.	Acc@50	Strict@50	$\Delta_{\mathrm{Acc}}$ (%)	$\Delta_{\mathrm{strict}}$ (%)
GPT-5	$54.2$	$22.0$	$42.9$	$50.0$	$43.2$	$17.7$	$20.3$	$19.5$
GPT-4o	$37.9$	$20.8$	$34.9$	$36.0$	$25.7$	$14.3$	$32.3$	$31.3$
Open-weight models
TrustMed-RL (GiGPO, 8B)	$\mathbf{37.1}_{\scriptscriptstyle\pm 1.1}$	$15.3_{\scriptscriptstyle\pm 1.0}$	$\mathbf{31.8}_{\scriptscriptstyle\pm 1.1}$	$\mathbf{34.6}_{\scriptscriptstyle\pm 1.1}$	$\mathbf{32.5}_{\scriptscriptstyle\pm 1.3}$	$\mathbf{13.5}_{\scriptscriptstyle\pm 0.6}$	$12.3$	$11.9$
TrustMed-RL (GRPO, 8B)	$34.5_{\scriptscriptstyle\pm 1.6}$	$13.6_{\scriptscriptstyle\pm 1.9}$	$28.9_{\scriptscriptstyle\pm 1.9}$	$32.1_{\scriptscriptstyle\pm 1.5}$	$30.8_{\scriptscriptstyle\pm 1.3}$	$12.2_{\scriptscriptstyle\pm 1.6}$	$\mathbf{10.7}$	$\mathbf{10.9}$

Appendix

Table 5: Diagnostic performance on 2,500 clean cases per evaluation run. All scores are percentages. SFT, GiGPO, and GRPO each use three training seeds, with three separately seeded evaluation passes per trained model (nine runs per method). Where reported, means and sample standard deviations (std) are computed across the three evaluation passes.

\Delta_{\mathrm{Acc}}=100(\mathrm{Acc}-\mathrm{Acc@50})/\mathrm{Acc}

and

\Delta_{\mathrm{strict}}=100(\mathrm{Acc}^{\mathrm{strict}}-\mathrm{Strict@50})/\mathrm{Acc}^{\mathrm{strict}}

measure relative accuracy reductions under the evidence-coverage requirement (lower is better). Non-diagnosis outcomes receive zero. Bold indicates the best result among open-weight models.

Proprietary (closed-source) models
Model	Order	Findings	Prod.	Workup	Search	Invalid	Modality Catch	Anatomy Catch	False Alarm
GPT-5	$72.4$	$62.8$	$8.5$	$12.8$	$63.2$	$0.0$	$97.8$	$88.7$	$4.8$
GPT-4o	$62.0$	$53.5$	$6.9$	$10.9$	$75.6$	$0.0$	$95.4$	$85.7$	$6.4$
Open-weight models
TrustMed-RL (GiGPO, 8B)	$83.9_{\scriptscriptstyle\pm 4.0}$	$73.2_{\scriptscriptstyle\pm 3.9}$	$8.5_{\scriptscriptstyle\pm 0.4}$	$17.8_{\scriptscriptstyle\pm 0.8}$	$93.0_{\scriptscriptstyle\pm 2.5}$	$4.0_{\scriptscriptstyle\pm 1.1}$	$82.2_{\scriptscriptstyle\pm 1.1}$	$52.6_{\scriptscriptstyle\pm 4.2}$	$2.2_{\scriptscriptstyle\pm 0.8}$
TrustMed-RL (GRPO, 8B)	$\mathbf{84.7}_{\scriptscriptstyle\pm 1.3}$	$\mathbf{73.9}_{\scriptscriptstyle\pm 1.1}$	$\mathbf{9.1}_{\scriptscriptstyle\pm 0.2}$	$18.3_{\scriptscriptstyle\pm 0.6}$	$89.3_{\scriptscriptstyle\pm 4.7}$	$2.6_{\scriptscriptstyle\pm 0.2}$	$85.4_{\scriptscriptstyle\pm 7.9}$	$60.6_{\scriptscriptstyle\pm 2.4}$	$2.4_{\scriptscriptstyle\pm 0.5}$

Appendix

Table 6: Agentic behavior for models in Table 5 . Order and Findings measure mean test-record ordering and findings coverage over 2,452 eligible clean cases. Findings includes laboratory, physiologic, imaging, and pathology reports, excluding raw images. Prod. counts actions yielding new history facts, images, or findings; Workup counts workup turns, both per clean episode. Search measures clean episodes with any search; Invalid measures invalid clean-episode turns. Modality and Anatomy Catch measure exposure-conditioned mismatch detection. False Alarm is the false-positive rate on matched clean-image gate-replay controls. All metrics except Prod. and Workup are percentages; — denotes unavailable results.

Model	MedQA	NEJM	MTMedDialog
HealthGPT-Pro-27B	$\mathbf{73.0}_{\scriptscriptstyle\pm 1.2}$	$\mathbf{35.6}_{\scriptscriptstyle\pm 3.4}$	$39.5_{\scriptscriptstyle\pm 48.9}$
Qwen3-VL-32B-Instruct	$65.3_{\scriptscriptstyle\pm 0.7}$	$31.4_{\scriptscriptstyle\pm 1.1}$	$28.3_{\scriptscriptstyle\pm 45.0}$
MedGemma-27B-it	$70.4_{\scriptscriptstyle\pm 1.0}$	$34.0_{\scriptscriptstyle\pm 1.5}$	$33.3_{\scriptscriptstyle\pm 47.1}$
Hulu-Med-32B	$55.1_{\scriptscriptstyle\pm 49.9}$	$22.2_{\scriptscriptstyle\pm 41.8}$	$29.1_{\scriptscriptstyle\pm 45.5}$
MedVLMThinker-32B-RL	$65.3_{\scriptscriptstyle\pm 1.9}$	$29.7_{\scriptscriptstyle\pm 4.0}$	$36.3_{\scriptscriptstyle\pm 48.1}$
Lingshu-32B	$56.2_{\scriptscriptstyle\pm 1.6}$	$24.1_{\scriptscriptstyle\pm 3.2}$	$31.3_{\scriptscriptstyle\pm 46.4}$

Appendix

Table 7: Out-of-distribution diagnostic accuracy (Acc, %). Subscripted

\pm

values denote standard deviations. Bold scores indicate the highest mean for each dataset.

	Solvable				Unsolvable
Model / variant	ACC	Strict	ACC@3	Strict@3	ACC	Strict	ACC@3	Strict@3	Reference	$\Delta$ solvable ACC [95% CI]
2B Base	2.27	0.60	6.21	1.65	0.44	0.09	1.07	0.27	-	-
2B + SFT	14.57	4.33	31.62	11.20	4.00	0.98	9.33	2.93	2B Base	+12.30 [11.28, 13.33]
8B Base	18.40	7.39	36.71	15.95	4.18	0.71	10.40	2.13	-	-
8B + SFT	27.89	10.40	48.94	21.65	7.64	2.04	17.33	5.07	8B Base	+9.49 [8.03, 10.92]
TrustMed-RL w/o step credit (GRPO)	39.98	17.77	60.00	32.28	11.02	3.91	21.87	8.53	8B SFT	+12.09 [10.59, 13.62]

Appendix

Table 8: Ablation study. Accuracy values are percentages; gains are percentage points relative to the named reference. ACC@3 and Strict@3 measure the proportion of cases correct in at least one of three evaluations under the corresponding criterion; each case contributes once. Base denotes Qwen3-VL Instruct. All RL variants use SFT initialization. Brackets show paired-case 95% intervals; the four SFT/RL gains also survive Bonferroni adjustment.

Task outcomes
Metric	Full	w/o priority sampling	w/o adaptive sampling	w/o adaptive sampling + horizon	w/o std	w/o $A^{S}$ (GRPO)
Overall task success (%)	38.63	33.45	29.64	32.40	34.39	38.72
Diagnostic success: answer (%)	28.01	20.83	18.53	21.84	19.90	22.61
Hard-band diagnostic success (%)	26.46	15.16	16.30	18.27	17.47	21.92
Groups with $\geq 1$ correct diagnosis (%)	68.12	58.37	60.25	67.23	61.57	60.98
Terminal reward	0.3174	0.2661	0.2391	0.2526	0.2797	0.3180

Appendix

Table 9: Late-training ablations over updates 217–256. Values are equally weighted means of available per-update statistics on each variant’s sampled training tasks, not held-out evaluation results. Percentage metrics use a 0–100 scale. Bold marks the highest outcome or lowest error mean; coverage and interaction/output lengths are descriptive.

Level	Weight	Definition
Full	$w_{\mathrm{F}}=1.00$	Identifies the same fully specified disease. Synonyms, abbreviations, spelling variants, and clinically equivalent terminology are accepted.
Core	$w_{\mathrm{C}}=0.85$	Identifies the same underlying disease but omits or misstates one or more clinically relevant qualifiers, such as subtype, site, etiology, pathology, severity, grade, stage, or complication.
Partial	$w_{\mathrm{P}}=0.50$	Correctly identifies an explicit constituent of the reference diagnosis, such as its etiologic agent, disease-defining manifestation, or complication, without identifying the core disease.
Null	$w_{\varnothing}=0$	Has no valid semantic correspondence with the reference. Shared anatomy, nonspecific symptoms, broad disease-family membership, or general clinical similarity alone do not constitute a match.

Appendix

Table 10: Semantic diagnosis-matching rubric used by the LLM judges.

	ORDO ( $N=100$ )	Physician-inclusive ( $N=139$ )
Judge	MAE $\downarrow$	Agree. (%) $\uparrow$	MAE $\downarrow$
GPT-4.1-mini	0.169 $\to$ 0.017	46.0 $\to$ 46.0	0.334 $\to$ 0.321
GPT-5-mini	0.098 $\to$ 0.020	51.1 $\to$ 51.1	0.269 $\to$ 0.262
GPT-4o	0.159 $\to$ 0.019	46.0 $\to$ 46.8	0.333 $\to$ 0.314
GPT-4o-mini	0.392 $\to$ 0.017	42.4 $\to$ 43.9	0.377 $\to$ 0.351
Claude Sonnet 5	0.067 $\to$ 0.037	55.4 $\to$ 53.2	0.249 $\to$ 0.257

Appendix

Table 11: Judge and ontology-floor ablations. Entries report without

\to

with the ontology floor. ORDO measures conformance to prespecified ontology-relation rewards; physician-inclusive evaluation measures agreement with a mixed human–model reference. Agreement denotes exact four-level label agreement. All settings use three passes.

Relation class	$n$	Prediction $\to$ reference construction	Target	Resampling unit
Synonym	20	Distinct ORDO synonym $\to$ preferred name	Full (1.00)	Pair
Narrower	20	Subtype $\to$ parent disorder	Core (0.85)	Hierarchy edge
Broader	20	Parent disorder $\to$ subtype	Core (0.85)	Hierarchy edge
Sibling subtype	10	Two non-ancestral subtypes of one disorder	Core (0.85)	Pair
Sibling group	10	Two disorders in one small ORDO group	None (0)	Pair
Unrelated	20	No shared clinical ancestor at any depth	None (0)	Pair

Appendix

Table 12: Construction of the ORDO-derived judge-conformance set. Relation classes map deterministically to expected rewards. Narrower and broader pairs are opposite orientations of the same 20 hierarchy edges and are resampled jointly, yielding 80 bootstrap clusters. The fixed class mixture has mean target reward

0.625

and does not represent deployment prevalence.

Judge	Synonym (1.00)	Narrower (0.85)	Broader (0.85)	Sibling subtype (0.85)	Sibling group (0.00)	Unrelated (0.00)
GPT-4.1-mini	0.878 $\to$ 1.000	0.680 $\to$ 0.850	0.680 $\to$ 0.850	0.085 $\to$ 0.850	0.000 $\to$ 0.170	0.000 $\to$ 0.000
GPT-5-mini	0.993 $\to$ 1.000	0.772 $\to$ 0.858	0.772 $\to$ 0.858	0.340 $\to$ 0.850	0.085 $\to$ 0.170	0.000 $\to$ 0.000
GPT-4o	0.935 $\to$ 1.000	0.730 $\to$ 0.858	0.722 $\to$ 0.850	0.000 $\to$ 0.850	0.085 $\to$ 0.170	0.000 $\to$ 0.000
GPT-4o-mini	0.485 $\to$ 1.000	0.340 $\to$ 0.850	0.340 $\to$ 0.850	0.000 $\to$ 0.850	0.000 $\to$ 0.170	0.000 $\to$ 0.000
Claude Sonnet 5	0.993 $\to$ 1.000	0.830 $\to$ 0.872	0.815 $\to$ 0.858	0.645 $\to$ 0.850	0.220 $\to$ 0.305	0.000 $\to$ 0.000
MedGemma-27B	0.887 $\to$ 1.000	0.762 $\to$ 0.858	0.797 $\to$ 0.858	0.370 $\to$ 0.850	0.135 $\to$ 0.220	0.000 $\to$ 0.000

Appendix

Table 13: Class-wise ontology-floor ablation on 100 ORDO pairs. All judges use the final strict-entity rubric and are evaluated in three passes. Each

a\to b

entry denotes without

\to

with the ontology floor. Class columns report mean rewards, with targets in parentheses; MAE is the mean absolute reward error over all 100 pairs. Bold marks only the lowest overall MAE, including ties. E1 uses MedGen, MONDO, and ORDO; E2 removes ORDO. Because ORDO supplies both the pairs and part of the floor, this is partly a self-consistency test.

Protocol	Synonym	Narrower	Broader	Sibling subtype	Sibling group	Unrelated	Overall	3/3 agreement
RareArena (2/1/0)	19/20	5/20	20/20	8/10	7/10	20/20	79/100 (78–80)	98/100
MeDxAgent (true/false)	19/20	13/20	16/20	10/10	10/10	20/20	88/100 (87–90)	93/100

Appendix

Table 14: Native-rule fidelity of our re-implementations of published judge protocols and descriptive mapping to our reward scale. Native fidelity compares each implementation with the labels prescribed by its own protocol; class counts use modal labels and the parenthesized overall ranges span the three repeats. Cross-rubric reward discrepancy maps native outputs to our rewards and is not a measure of protocol accuracy.

Figure 3: Doctor prompt for history-taking.

Figure 4: Doctor prompt for diagnostic workup.

Figure 5: Doctor prompt for evidence-integrity verification.

Figure 7: Simulated-patient prompt for fact-grounded responses.

Criterion	Mean $\pm$ SD	Scores 4–5, $n$ (%)	Scores 1–2, $n$ (%)
C1: Findings interpretation	$4.33\pm 0.91$	158 (79.0)	8 (4.0)
C2: Evidence use and revision	$4.06\pm 0.93$	144 (72.0)	11 (5.5)
C3: Evidential grounding	$4.39\pm 0.92$	166 (83.0)	12 (6.0)

Appendix

Table 15: Physician assessment of 200 automatically accepted diagnostic trajectories. Scores range from 1 to 5. We report the mean and sample standard deviation across trajectories, together with counts and percentages in the indicated score ranges.

Panel of Figure 9 — Figure 9: Example of a multi-turn clinical diagnostic trajectory in our interactive environment.

Explore similar work

OpenMedReason: Scientific Reasoning Supervision for Medical Vision-Language Models
Jun 10, 2026Negin Baghbanzadeh, Pritam Sarkar, Michael Colacci +6Medical VQA Medical VLMs
York University · Vector Institute · University of British Columbia +5
MedAction: Towards Active Multi-turn Clinical Diagnostic LLMs
May 8, 2026Hsin-Ling Hsu, Zizheng Wang, Donghua Zhang +9Healthcare Medical Diagnosis
1National Chengchi University · 2Georgetown University · University of Michigan +3
Med-R2: An Adversarial Benchmark for Evidence-Grounded Reasoning in Medical VLMs
May 23, 2026Wen Ma, Fucheng Niu, Zhiting Fan +3Adversarial Attacks on VLMs Healthcare
Zhejiang University, Zhejiang, China · Guangdong Institute of Intelligence Science and Technology

Jun 10, 2026, cs.CVJ/K move · Enter open · S save

OpenMedReason: Scientific Reasoning Supervision for Medical Vision-Language Models

Negin Baghbanzadeh, Pritam Sarkar, Michael Colacci, Abeer Badawi+5

York University · Vector Institute · University of British Columbia+5

Medical VQA Medical VLMs VLM Reasoning Visually Grounded Reasoning Medical Image Benchmarks

Open paper AskPDF arXiv

High-stakes clinical use of large vision-language models (LVLMs) requires reasoning that is grounded in visual evidence and clinical knowledge, not just correct final answers. We introduce OpenMedReason, a large-scale, open multimodal medical reasoning corpus comprising approximately 450K image-question-answer instances whose reasoning traces are primarily derived from curated biomedical, human-authored scientific articles. OpenMedReason provides high-fidelity supervision beyond synthetic chains of thought, covering diverse medical domain vision modalities such as radiological scans, microscopic images, visible light photographs, charts, and others. We complement it with OpenMedReason-Bench, a held-out benchmark that allows fine-grained evaluation of LVLMs along three complementary axes of capability, including perception, medical knowledge, and rationale, enabling diagnostic evaluation beyond final-answer accuracy. OpenMedReason is a rich training resource that exhibits its effectiveness in both supervised fine-tuning (SFT) and reinforcement-based alignment. Training with OpenMedReason yields a 20% average improvement in VQA accuracy over the base model and achieves performance within 4.2% of the strongest comparable-scale medical LVLMs. Fine-grained performance analysis confirms that the gains are not concentrated in any single axis: OpenMedReason improves perception, medical knowledge, and rationale jointly, and its reasoning traces are preferred over those of the base model in 86.1% of pairwise comparisons. We release the code and dataset at huggingface.co/datasets/neginb/OpenMedReason.

Abstract

Explore similar work

OpenMedReason: Scientific Reasoning Supervision for Medical Vision-Language Models

MedAction: Towards Active Multi-turn Clinical Diagnostic LLMs

Med-R2: An Adversarial Benchmark for Evidence-Grounded Reasoning in Medical VLMs