ClinVerse: A Realistic Multi-Agent Diagnostic Environment for Training and Evaluating Medical Agents

Sirry Chen; Fangke Chen; Xuanwen Ding; Wei Chen; Zhongyu Wei

cs.AIOct 26, 2026

ClinVerse: A Realistic Multi-Agent Diagnostic Environment for Training and Evaluating Medical Agents

Authors: Sirry Chen, Fangke Chen, Xuanwen Ding, Wei Chen, Zhongyu Wei

Organizations: Fudan University, Shanghai, China · Shanghai Innovation Institute, Shanghai, China · Zhejiang University, Hangzhou, China · Huazhong University of Science and Technology, Wuhan, China ·

Abstract

Disease diagnosis is a key application of large language model (LLM) agents in healthcare. However, existing medical agents are primarily trained and evaluated with static question-answering data, providing limited opportunities to learn active information acquisition, examination decision-making, and dynamic clinical interaction. To address this limitation, we present ClinVerse, an interactive, real-data-grounded multi-agent environment designed for training and evaluating medical diagnostic agents. Built on ClinPatient, a scalable dataset comprising 150k diagnostic cases derived from real-world electronic health records, ClinVerse supports coordinated interactions among six specialized clinical roles. Specifically, a doctor agent starts from an initial patient complaint, progressively gathers clinical evidence through patient interviews and examination requests, and ultimately produces a diagnosis. For each interaction trajectory, a rule-based scoring framework quantifies information acquisition, examination ordering, and diagnostic accuracy. These trajectory-level scores serve as reward signals for agent training while also enabling scalable evaluation. In experiments, we evaluate various general-purpose and medical-domain models and show that nine self-evolution methods across three categories consistently improve agents' diagnostic capabilities within ClinVerse. We further analyze the interplay among model capability, environment feedback, and self-evolution methods.

Figures & tables

Work	Venue	Cases	Data Source	Roles	Scalable	Real Data	SME
Static Question-answering
MedQA ( Jin et al., 2020 )	arXiv 2020	61k	Medical licensing exams	—	✗	✔	✗
CMExam ( Liu et al., 2023 )	NeurIPS 2023	68k	Chinese medical licensing exam	—	✗	✔	✗
MedXpertQA ( Zuo et al., 2025 )	ICML 2025	4,460	Medical examinations	—	✗	✔	✗
EHRBench ( Xie et al., 2026 )	KDD 2026	960k	MIMIC-III, MIMIC-IV, etc.	—	✔	✔	✗
Interactive Clinical Environments

Table 1. Comparison of static medical QA and interactive clinical environments. ClinVerse is grounded in 150k scalable real-world patient data, and supports predicting missing examination results to enable flexible interactions of 6 roles. Role icons denote medical agent, laboratory physician, nurse, patient, judge, and environment controller. SME means supporting missing examinations.

Figure 1. An overview of ClinVerse . Built upon 150k EHR-derived dataset ClinPatient , a dual-layer patient simulator ensures controlled information disclosure, while the laboratory physician retrieves recorded examination results or predicts missing ones using ClinLab . ClinVerse integrates six specialized roles to support multi-turn open-ended clinical interaction. The resulting interaction trajectories are evaluated using three-dimensional rule-based metrics and further utilized for three levels of agent self-evolution strategies.

Figure 2. Dataset Statistics of ClinPatient .

Doctor Agent	Info Collection Rate (%)	Test Order Coverage (%)	Abnormal Test Coverage (%)	Diagnosis Category Accuracy (%)	Diagnosis Subtype Accuracy (%)
L0: Static Agents
Lingshu-32B	$28.82\pm 2.69$	$0.58\pm 0.18$	$1.32\pm 0.29$	$28.20\pm 2.17$	$11.60\pm 1.95$
Baichuan-M2-32B	$26.90\pm 0.72$	$56.30\pm 1.92$	$65.60\pm 2.13$	$29.41\pm 3.40$	$19.22\pm 4.02$
HealthGPT-Pro-4B	$12.79\pm 1.87$	$22.04\pm 1.18$	$24.24\pm 1.60$	$25.67\pm 3.06$	$9.00\pm 1.73$
HealthGPT-Pro-8B	$10.62\pm 1.41$	$33.90\pm 2.73$	$36.56\pm 3.43$	$24.00\pm 8.34$	$11.60\pm 5.46$
GPT-4o-mini	$27.49\pm 0.41$	$47.45\pm 2.42$	$60.98\pm 2.64$	$32.20\pm 1.10$	$19.20\pm 2.39$

Table 2. Performance comparison of different doctor models and training methods. Results are grouped by four levels and reported as mean

\pm

standard deviation (%). Methods prefixed with “+” are applied to Qwen3-32B; bold denotes the best result among these adaptation methods. GPT-5.6-Sol with full information is included as a performance reference.

Figure 3. (a) Average information collection rate across dialogue turns. Information is acquired rapidly early in the consultation and gradually plateaus; (b) Average agreement between turn-level and final diagnoses, which increases as more evidence is collected; (c) Distribution of the four action types across turns, showing a general progression from patient inquiry and examination to testing and final diagnosis.

Traj Model	Traj Num	Doctor Model	IC	TOC	ATC	DCA	DSA
Qwen3-32B	5000	GPT-4o-mini	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 0.30}}$	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 6.16}}$	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 11.41}}$	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 8.20}}$	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 4.15}}$
	5000	GPT-5.6-Luna	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 1.86}}$	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 1.94}}$	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 0.65}}$	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 1.31}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 1.49}}$
	5000	Qwen3-8B	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 0.86}}$	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 2.55}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 2.65}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 0.97}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 3.77}}$
	5000	Qwen3-32B	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 0.40}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 2.18}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 6.10}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 5.36}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 4.37}}$
	500	Qwen3-32B	${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 0.55}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 1.60}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 3.79}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 0.52}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 4.44}}$
GPT-5.6-Luna	500	GPT-4o-mini	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 4.24}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 0.86}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 1.32}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 1.80}}$	${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 11.80}}$

Table 3. Performance variation of models training with MemRL using trajectories from different models. All results are reported in percentage (%) and the five metrics are abbreviated forms of these reported in Table 2 . Cross-model adaptation depends strongly on the source–target capability relationship, with stronger-model trajectories often providing greater benefits despite smaller data volumes.

Doctor Agent	MVME	Recall	Precision	F1
Qwen3-32B	45.67	18.81	30.42	23.25
+ MemRL-offline	47.40 ↑1.73	20.41 ↑1.60	32.78 ↑2.36	25.16 ↑1.91
+ SkillRL	44.20 ↓1.47	19.69 ↑0.88	31.28 ↑0.86	24.17 ↑0.92
+ MedAgent-Zero	45.73 ↑0.06	18.93 ↑0.12	31.47 ↑1.05	23.55 ↑0.30
+ Trace2Skill	45.13 ↓0.54	22.94 ↑4.13	35.89 ↑5.47	27.99 ↑4.74
+ Dynamic Cheatsheet	39.45 ↓6.22	26.12 ↑7.31	36.85 ↑6.43	30.57 ↑7.32

Table 4. Generalization performance of Qwen3-32B adapted with different training methods using interaction trajectories from ClinVerse and evaluated on AIHospital.

Figure 4. Information collection and diagnostic reasoning capabilities across 3 levels of evolution methods and backbone models. Information collection capability is calculated as the average of ICR, TOC, and ATC, while diagnostic accuracy is calculated as the average of DCA and DSA. Greater information acquisition does not necessarily yield better diagnostic accuracy, highlighting the importance of how medical agents select and utilize clinical evidence.

Setting	Arbitrary Disclosure Rate	Test-related Disclosure Rate	Refusal Rate
Single&Neutral	47.2%	87.0%	10.0%
Dual-layer&Neutral	36.8%	66.0%	19.7%
Single&Self-check	25.5%	67.0%	62.8%
Dual-layer&Self-check	5.7%	13.5%	84.7%

Table 5. Comparison of information disclosure and refusal behaviors under different patient simulation settings.

Model	MAPE	Median MAPE	MAE	NormRefErr	Anomaly Recall
Qwen3.5-2B	57.36	27.17	77.21	777.23	30.68
Qwen3.5-27B	35.33	11.54	22.45	2.01	59.69
DeepSeek-V4-Flash	35.47	12.50	61.63	1.92	63.95
GPT-4o-mini	44.52	18.43	57.28	4.73	69.04
ClinLab-2B	19.46	3.70	14.62	0.36	76.92

Table 6. Performance comparison across different models as physiological-state predictor. Our trained model ClinLab accurately predicts examination outcomes while achieving high recall for abnormal findings.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Info. Collection	Test Order	Abnormal Test	Diag. Category	Diag. Subtype
Qwen3-8B	$\mathbf{14.3\pm 1.1}$	$40.7\pm 1.8$	$48.4\pm 2.4$	$23.2\pm 2.4$	$7.8\pm 2.8$
+ SkillRL-GRPO	$6.7\pm 0.2$	$\mathbf{45.2\pm 1.5}$	$\mathbf{57.6\pm 1.7}$	$28.6\pm 2.1$	$16.6\pm 1.1$
+ Full Info	–	–	–	$\mathbf{32.6}$	$\mathbf{17.5}$

Appendix

Table 7. Performance of Qwen3-8B before and after SkillRL-GRPO training. Interactive results are reported as mean

\pm

standard deviation over five runs of 100 cases. The Full Info reference is evaluated on 200 cases and does not involve information collection or test ordering.

Figure 5. Comparison of Information Collection Rate of Different Training Methods with Qwen3-32B

Figure 6. Comparison of Action Taken Ratio of Different Training Methods with Qwen3-32B

Doctor Model	Diagnosis Category Accuracy (%)	Diagnosis Subtype Accuracy (%)
GPT-5.6-Luna	48.00	38.00
GPT-5.6-Sol	54.00	45.00
Qwen3-8B	32.60	17.50
Qwen3-32B	35.00	21.60
HealthGPT-Pro-4B	32.50	9.50
HealthGPT-Pro-8B	32.40	17.80

Appendix

Table 8. Diagnostic performance of different models under the full-information setting.

Doctor Model	Method	IC	TOC	ATC	DCA	DSA
Qwen3-32B	Static	13.96	40.84	48.06	26.04	12.43
	MemRL	14.36	43.02	54.16	31.40	16.80
	Dynamic Cheatsheet	11.48	49.33	57.45	33.40	20.60
	Self-Taught Optimizer	12.89	40.32	50.85	29.40	17.60
GPT-4o-mini	Static	27.49	47.45	60.98	32.20	19.20
	MemRL	28.06	50.21	64.35	38.00	26.00

Appendix

Table 9. Performance of three doctor models under static inference and three adaptation methods. IC, TOC, ATC, DCA, and DSA denote Info Collection, Test Order Coverage, Abnormal Test Coverage, Diagnosis Category Accuracy, and Diagnosis Subtype Accuracy, respectively.

Method	IC	TOC	ATC	DCA	DSA
MemRL	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 2.24}$	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 4.80}$	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 4.71}$	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 1.60}$	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 1.00}$
SkillRL	${\color[rgb]{0.8945,0.2109,0.2109}\uparrow 3.23}$	${\color[rgb]{0.8945,0.2109,0.2109}\uparrow 2.12}$	${\color[rgb]{0.8945,0.2109,0.2109}\uparrow 3.49}$	${\color[rgb]{0.8945,0.2109,0.2109}\uparrow 0.40}$	${\color[rgb]{0.8945,0.2109,0.2109}\uparrow 1.20}$
MedAgent-Zero	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 5.71}$	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 15.89}$	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 18.29}$	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 7.00}$	${\color[rgb]{0.1328,0.5469,0.1328}\downarrow 4.60}$

Appendix

Table 10. Changes relative to Baseline for Qwen3-32B agents adapted in AIHospital and evaluated in ClinVerse . Metric abbreviations follow Table 2 . Upward and downward arrows indicate increases and decreases, respectively.

Doctor Model	Diagnosis Subtype Accuracy (%)
GPT-4o-mini	51 ${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 31.80}}$
GPT-5.6-Luna	73 ${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 40.11}}$
Qwen3-32B	45 ${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 32.57}}$
Lingshu-32B	49 ${{\color[rgb]{0.8945,0.2109,0.2109}\uparrow 37.40}}$
Baichuan-M2-32B	14 ${{\color[rgb]{0.1328,0.5469,0.1328}\downarrow 5.22}}$

Appendix

Table 11. Diagnosis subtype accuracy under full information with the ground-truth ICD-10 block provided as a hard constraint. The model must predict the correct three-character category within the given block.

Replacement Ratio	DCA	DSA
0%	26.04 $\pm$ 2.69	12.43 $\pm$ 1.12
10%	26.41 $\pm$ 2.30	13.22 $\pm$ 1.48
20%	28.27 $\pm$ 1.92	14.61 $\pm$ 1.52

Appendix

Table 12. Diagnostic performance under different proportions of ClinLab-predicted examination values.

Setting	DCA	DSA
w/ ClinLab	35.58	31.73
w/o ClinLab	39.17	31.91

Appendix

Table 13. Diagnostic performance with and without ClinLab.

Subgroup	ICR	TOC	ATC	Category Acc.	Subtype Acc.
Male	13.2	35.6	46.9	42.0	24.0
Female	13.2	34.2	45.8	24.0	15.0
Age 18–44	15.2	39.2	48.5	33.0	18.0
Age 45–64	12.0	37.9	44.9	31.0	18.0
Age 65–74	12.9	38.6	45.7	30.0	18.0
Age 75+	11.1	41.0	51.2	33.0	16.0

Appendix

Table 14. Subgroup performance of the fixed Qwen3-32B doctor agent. All values are percentages over completed evaluations. ICR, TOC, and ATC denote information collection rate, test-order coverage, and abnormal-test coverage, respectively. Category and subtype accuracy measure agreement at the ICD-10 block and three-character category levels. The urinary-system cohort includes N10–N16 and N30–N39.

Cohort	Case ID	Reference	Prediction	ICR	TOC	ATC	Category Acc.	Subtype Acc.
Heart failure	10199560_22434262	I50.23	I50.1	6.4	43.1	100.0	100.0	100.0
Sepsis	10651989_24805766	A41.9	K80.3	19.5	67.4	100.0	0.0	0.0
Pneumonia	10333909_27924784	J18.9	I50	32.7	91.7	100.0	0.0	0.0

Appendix

Table 15. Illustrative diagnostic trajectories across disease cohorts. All scores are percentages. Case identifiers refer to the frozen evaluation records. Reference and predicted diagnoses retain their original ICD-10 codes, while diagnostic accuracy is evaluated at the coarser levels defined in Section 2.4 .

Success: iterative evidence acquisition supports the correct ICD-10 category
Presentation	A 57-year-old patient with hypertension, hyperlipidemia, diabetes, peripheral neuropathy, and current smoking presented with worsening positional dizziness, episodic severe hypertension, and bilateral hand paresthesia. Examination showed a left cervical bruit and subtle right-sided neurological abnormalities.
Diagnosis	Agent output: cerebral infarction , coded as I63. Reference: occlusion/stenosis of a carotid artery with cerebral infarction (I63.239).
Metric scores	$R_{\mathrm{info}}=$ 0.159 $R_{\mathrm{test}}=$ 0.409 $R_{\mathrm{abnormal}}=$ 0.111 $A_{\mathrm{cat}}=$ 1 $A_{\mathrm{sub}}=$ 1
Failure: high evidence coverage still leads to a symptom-level answer
Presentation	A 73-year-old patient with diabetes, diastolic heart failure, hypertension, possible coronary artery disease, and chronic kidney disease presented with progressive dyspnea and dry cough. Symptoms worsened with exertion and recumbency, but there was no chest pain.
Diagnosis	Agent output: dyspnea and dry cough , mapped to R06.0. Reference: subendocardial infarction/NSTEMI (I21.4).

Appendix

Table 16. Representative successful and failed Qwen3-32B trajectories.

Turn	Information route	Actual interaction
Init.	EC $\rightarrow$ D	EC: Revealed a 57-year-old patient with hypertension, hyperlipidemia, diabetes with peripheral neuropathy, and current smoking. Admission examination included a left cervical bruit, subtle right-sided weakness and clumsiness, and an unsteady gait.
1	D $\leftrightarrow$ OPS	D: “Describe your main symptoms, their onset and duration, and what worsens or relieves them.” OPS: “I have recurrent dizziness, especially when standing. My blood pressure sometimes reaches 190–210, and both hands feel numb and tingly.”
2	D $\leftrightarrow$ OPS	D: “When did the dizziness begin, and was the onset sudden or gradual?” OPS: “It began about one month ago, initially only occasionally, but has worsened over the past few days and is most noticeable whenever I stand.”
3	D $\leftrightarrow$ OPS	D: “Does the dizziness occur with nausea, blurred vision, palpitations, or sweating?” OPS: “I sometimes have transient blurred vision, but no nausea, palpitations, or sweating.”
4	D $\rightarrow$ LP LP $\leftrightarrow$ LPS LP $\rightarrow$ D	D: Requested vital-sign and neurological examinations to evaluate dizziness, blood-pressure fluctuation, and paresthesia. LP (from LPS): BP 115/74, left cervical bruit, distal sensory loss, slow/clumsy right-hand movements, and an unsteady gait.
5	D $\rightarrow$ LP LP $\leftrightarrow$ LPS/ClinLab LP $\rightarrow$ D	D: Ordered a metabolic panel, cardiac markers, and inflammatory markers. LP (from LPS/ClinLab): Glucose 214 mg/dL; creatinine 0.8 mg/dL; sodium 138 mEq/L; potassium 4.3 mEq/L; troponin T $<0.01$ ng/mL; CK-MB 3.0 ng/mL. Additional requested items absent from the source record were generated by ClinLab .

Appendix

Table 17. Turn-by-turn successful diagnostic trajectory for a de-identified case (10006457_27894366). D denotes Doctor, OPS Observable Patient Simulator, LPS Latent Patient Simulator, LP Laboratory Physician, EC Environment Controller, and DJ Diagnostic Judge. Patient responses are filtered by the self-check mechanism before being returned to D, while clinical evidence is provided through LP with support from LPS and ClinLab when needed. Blue, orange, and green rows represent patient interaction, evidence acquisition, and diagnosis submission, respectively.

Medical AI Agents

See all 38 papers

MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments
May 8, 2026Yicheng Gao, Xiaolin Zhou, Yahan Li +2Healthcare Medical Diagnosis
University of Southern California · Arizona State University
MeDxAgent: Multi-Agent Consultation for Interactive Medical Diagnosis
Jun 2, 2026Akshat Sanghvi, Naren Akash, Raza Imam +2Multi-Agent LLM Systems Medical Diagnosis
Microsoft Research India · Institute of AI and Innovation, AIG Hospitals · Mohamed bin Zayed University of Artificial Intelligence
ClinicalAgents: Multi-Agent Orchestration for Clinical Decision Making with Dual-Memory
Mar 27, 2026Zhuohan Ge, Haoyang Li, Yubo Wang +3Clinical Decision-Making Multi-Agent LLM Systems
The Hong Kong Polytechnic University, Hong Kong SAR, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China

May 8, 2026J/K move · Enter open · S save

MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments

Yicheng Gao, Xiaolin Zhou, Yahan Li, Yue Zhao+1

University of Southern California · Arizona State University

Healthcare Medical Diagnosis LLM Agent Training RL for LLM Agents LLM Agent Reliability Medical AI Agents

Open paper AskPDF arXiv

Real-world clinical diagnosis is a complex process in which the doctor is required to obtain information from both interaction with the patient and conducting medical exams. Additionally, the doctor needs to adapt to different patient personas, as well as noisy and incomplete information that can happen at any time during the process. However, existing benchmarks for medical LLMs and methods for automatic diagnosis largely simplify this process by reducing it to single-turn question answering, noise-free conversations, or sequential exam making, etc., ignoring the interactive and uncertain nature of clinical diagnosis. In this paper, we aim to address this gap by formalizing clinical diagnosis as a Partially Observable Markov Decision Process (POMDP) with three action types: questioning the patient, ordering medical exams as tool calls, and issuing a diagnosis. We also introduce a systematic noise model comprising seven patient noise types and three exam noise types. Using our proposed environment, we train an effective diagnosis agent, MedExAgent, through a two-stage pipeline that first performs supervised finetuning on synthetic conversations structured after the Calgary-Cambridge model for clinical interviews, and then applies DAPO to optimize a composite reward capturing diagnostic accuracy, tool call quality, and exam cost including financial cost and patient discomfort. Through extensive experiments and ablation studies, we demonstrate that MedExAgent achieves diagnostic performance comparable to larger models while maintaining cost-efficient examination strategies.

Abstract

Medical AI Agents

Explore similar work

MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments

MeDxAgent: Multi-Agent Consultation for Interactive Medical Diagnosis

ClinicalAgents: Multi-Agent Orchestration for Clinical Decision Making with Dual-Memory