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
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
Model
Acc
Acc strict
Prec.
Rec.
Acc@50
Strict@50
ΔAcc (%)
Δstrict (%)
Proprietary (closed-source) models
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)
37.1±1.1
15.3±1.0
31.8±1.1
34.6±1.1
32.5±1.3
13.5±0.6
12.3
11.9
TrustMed-RL (GRPO, 8B)
34.5±1.6
13.6±1.9
28.9±1.9
32.1±1.5
30.8±1.3
12.2±1.6
10.7
10.9
Table 1: Diagnostic performance on 2,500 clean cases. Scores are percentages; ± 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 .
Model
Order
Findings
Prod.
Workup
Search
Invalid
Modality Catch
Anatomy Catch
False Alarm
Proprietary (closed-source) models
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±4.0
73.2±3.9
8.5±0.4
17.8±0.8
93.0±2.5
4.0±1.1
82.2±1.1
52.6±4.2
2.2±0.8
TrustMed-RL (GRPO, 8B)
84.7±1.3
73.9±1.1
9.1±0.2
18.3±0.6
89.3±4.7
2.6±0.2
85.4±7.9
60.6±2.4
2.4±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
Δ 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 st and releases a multimodal observation ot to the physician. The physician selects an action at using its observable interaction context ht , and the environment updates its state through P . The available actions depend on the current interaction phase φt , remaining budgets bt , and evidence and integrity states. The patient simulator and clinical tools are environment components; only the physician policy πθ is optimized. R(τ) 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 DSFT ; the resulting SFT policy initializes πθ and serves as the frozen KL reference πref . (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
Ri
Diagnosis, subject to the overrides below
Ridx
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.
Model
Acc
Acc strict
Prec.
Rec.
Acc@50
Strict@50
ΔAcc (%)
Δstrict (%)
Proprietary (closed-source) models
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)
37.1±1.1
15.3±1.0
31.8±1.1
34.6±1.1
32.5±1.3
13.5±0.6
12.3
11.9
TrustMed-RL (GRPO, 8B)
34.5±1.6
13.6±1.9
28.9±1.9
32.1±1.5
30.8±1.3
12.2±1.6
10.7
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. ΔAcc=100(Acc−Acc@50)/Acc and Δstrict=100(Accstrict−Strict@50)/Accstrict 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.
Model
Order
Findings
Prod.
Workup
Search
Invalid
Modality Catch
Anatomy Catch
False Alarm
Proprietary (closed-source) models
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±4.0
73.2±3.9
8.5±0.4
17.8±0.8
93.0±2.5
4.0±1.1
82.2±1.1
52.6±4.2
2.2±0.8
TrustMed-RL (GRPO, 8B)
84.7±1.3
73.9±1.1
9.1±0.2
18.3±0.6
89.3±4.7
2.6±0.2
85.4±7.9
60.6±2.4
2.4±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
73.0±1.2
35.6±3.4
39.5±48.9
Qwen3-VL-32B-Instruct
65.3±0.7
31.4±1.1
28.3±45.0
MedGemma-27B-it
70.4±1.0
34.0±1.5
33.3±47.1
Hulu-Med-32B
55.1±49.9
22.2±41.8
29.1±45.5
MedVLMThinker-32B-RL
65.3±1.9
29.7±4.0
36.3±48.1
Lingshu-32B
56.2±1.6
24.1±3.2
31.3±46.4
Appendix
Table 7: Out-of-distribution diagnostic accuracy (Acc, %). Subscripted ± 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
Δ 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.
Metric
Full
w/o priority sampling
w/o adaptive sampling
w/o adaptive sampling + horizon
w/o std
w/o AS (GRPO)
Task outcomes
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 ≥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
wF=1.00
Identifies the same fully specified disease. Synonyms, abbreviations, spelling variants, and clinically equivalent terminology are accepted.
Core
wC=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
wP=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∅=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 ↓
Agree. (%) ↑
MAE ↓
GPT-4.1-mini
0.169 → 0.017
46.0 → 46.0
0.334 → 0.321
GPT-5-mini
0.098 → 0.020
51.1 → 51.1
0.269 → 0.262
GPT-4o
0.159 → 0.019
46.0 → 46.8
0.333 → 0.314
GPT-4o-mini
0.392 → 0.017
42.4 → 43.9
0.377 → 0.351
Claude Sonnet 5
0.067 → 0.037
55.4 → 53.2
0.249 → 0.257
Appendix
Table 11: Judge and ontology-floor ablations. Entries report without → 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 → reference construction
Target
Resampling unit
Synonym
20
Distinct ORDO synonym → preferred name
Full (1.00)
Pair
Narrower
20
Subtype → parent disorder
Core (0.85)
Hierarchy edge
Broader
20
Parent disorder → 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 → 1.000
0.680 → 0.850
0.680 → 0.850
0.085 → 0.850
0.000 → 0.170
0.000 → 0.000
GPT-5-mini
0.993 → 1.000
0.772 → 0.858
0.772 → 0.858
0.340 → 0.850
0.085 → 0.170
0.000 → 0.000
GPT-4o
0.935 → 1.000
0.730 → 0.858
0.722 → 0.850
0.000 → 0.850
0.085 → 0.170
0.000 → 0.000
GPT-4o-mini
0.485 → 1.000
0.340 → 0.850
0.340 → 0.850
0.000 → 0.850
0.000 → 0.170
0.000 → 0.000
Claude Sonnet 5
0.993 → 1.000
0.830 → 0.872
0.815 → 0.858
0.645 → 0.850
0.220 → 0.305
0.000 → 0.000
MedGemma-27B
0.887 → 1.000
0.762 → 0.858
0.797 → 0.858
0.370 → 0.850
0.135 → 0.220
0.000 → 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→b entry denotes without → 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 6: Doctor prompt for Conclusion
Figure 7: Simulated-patient prompt for fact-grounded responses.
Figure 8: Evaluator prompt.
Criterion
Mean ± SD
Scores 4–5, n (%)
Scores 1–2, n (%)
C1: Findings interpretation
4.33±0.91
158 (79.0)
8 (4.0)
C2: Evidence use and revision
4.06±0.93
144 (72.0)
11 (5.5)
C3: Evidential grounding
4.39±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.
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.
Negin Baghbanzadeh, Pritam Sarkar, Michael Colacci +6
York University · Vector Institute · University of British Columbia +5
Most existing LLM diagnoses are evaluated on static, single-turn settings where complete patient information is provided upfront, an oversimplification of real clinical practice. We study active diagnosis: the real-life clinical process of starting from initial observation, ordering tests, interpreting results, and updating a differential diagnosis across multiple turns. Through systematic analysis, we identify three recurring failure modes in current LLMs: ungrounded test ordering, unreliable diagnostic update, and degraded multi-turn coherence. Together, these failures reveal a core deficit: existing medical training data teaches models to reason from complete information but not to act under evolving, partial evidence. To address this gap, we introduce MedAction, a tree-structured distillation pipeline that synthesizes diverse and high-quality multi-turn diagnostic trajectories via LLM-environment interaction. We propose two knowledge-graph-grounded metrics to filter trajectory quality: Disease Trajectory Consistency (DTC), which tracks whether the model's hypothesis converges toward the correct diagnosis, and Reasoning-Action Consistency (RAC), which verifies that belief updates are driven by gathered evidence. Using this pipeline, we construct MedAction-32K, a dataset of 32,681 trajectories from 2,896 PMC cases. Fine-tuning an 8B model on MedAction-32K achieves state-of-the-art performance among open-source models on both MedR-Bench and our curated MedAction-300-Hard benchmark, pushing the edge for open-source medical LLMs.
Hsin-Ling Hsu, Zizheng Wang, Donghua Zhang +9
1National Chengchi University · 2Georgetown University · University of Michigan +3
Vision-language models have demonstrated impressive capabilities in general medical visual question answering, yet due to limited interpretability, it remains unclear whether their predictions reflect evidence-grounded clinical reasoning or reliance on spurious priors. We introduce Med-R2 Bench, a hierarchical benchmark aligned with the clinical workflow to evaluate adversarial robustness with visual grounding. We design stepwise QA tasks to assess whether reasoning chains are strictly grounded in visual evidence across the four clinical stages, and employ adversarial perturbations to test robustness against misleading cues. Med-R2 comprises 42,432 images, 31 task categories, and 110,406 QA pairs. Evaluation across 14 VLMs reveals a sequential performance degradation along the four-stage clinical workflow. Adversarial experiments show that models rely heavily on correct prompts to guess answers. Even when provided with explicit visual cues, the models struggle to accurately align textual descriptions. Finally, we demonstrate stepwise fine-tuning using our hierarchical data significantly improves reasoning robustness, highlighting its potential to drive future improvements in evidence-based medical AI.
Wen Ma, Fucheng Niu, Zhiting Fan +3
Zhejiang University, Zhejiang, China · Guangdong Institute of Intelligence Science and Technology