MedEvo: Medical Reasoning Evolution through Knowledge Navigation

Xiaoliang Wang; Zhuo Chen; Chuanyang Zheng; Quanbin Wang; Jiangwei Lao; Yingao Liu; Jiajia Liu; Jinming Cao

cs.AIOct 13, 2026

MedEvo: Medical Reasoning Evolution through Knowledge Navigation

Authors: Xiaoliang Wang, Zhuo Chen, Chuanyang Zheng, Quanbin Wang, Jiangwei Lao, Yingao Liu, Jiajia Liu, Jinming Cao,

Organizations: Zhejiang University · Ant Group · Shanghai Artificial Intelligence Laboratory · Shanghai Jiao Tong University

Abstract

Medical multimodal reasoning models are typically trained on teacher-generated rationales, discarding the valid reasoning in students' own failed trajectories, while general-domain correction methods offer little guidance for errors in image reading, anatomical context, or differential diagnosis. We propose MedEvo, which uses medical knowledge, namely a tag taxonomy and error-type diagnosis, to navigate both trajectory repair and data selection. For each failed rollout, a gold-blind auditor locates the first consequential error and provides a minimal, answer-free hint; the same student retains the valid prefix and regenerates the erroneous suffix. After each round, MedEvo identifies capability gaps across medical tags on a held-out feedback set and prioritizes related questions for new rollouts and repairs. With supervised fine-tuning (SFT) alone, MedEvo raises Qwen3.5-9B mean accuracy across seven benchmarks from 65.57% to 69.12% (47.40% to 54.50% on the multimodal track of MedXpertQA (MedXpertQA-MM)), surpassing teacher-distillation SFT (66.93%). It also raises Qwen3.5-35B and Qwen3.5-397B mean accuracy from 67.91% to 73.04% and from 71.83% to 74.95%, respectively. Our analyses show that failed trajectories contain reasoning worth preserving, that repair success does not imply learning value, and that independent feedback is key to sustained improvement.

Figures & tables

Figure 1: Medical multimodal reasoning across model scales. Mean accuracy on seven benchmarks. MedEvo improves 9B/35B/397B Qwen3.5 models. B: billion parameters; Hindsight Hint Distillation (HHD) [ 57 ] .

Figure 2: Local trajectory repair. A minimal, answer-blind hint guides the student to recover from a failed reasoning step while retaining its supported prefix. Localization and admission checks are detailed in § 3.2.1 and § 3.2.4 .

Figure 3: Medical multimodal data. Top: modality and difficulty; bottom: examples. Electrocardiogram (ECG); magnetic resonance cholangiopancreatography (MRCP); positron emission tomography/computed tomography (PET/CT); magnetic resonance imaging (MRI). Examples include both anatomical images and physiological recordings.

Figure 4: Knowledge Navigation in MedEvo. (a) Medical tags organize questions. (b) The current student generates fresh rollouts. (c) Knowledge-guided trajectory navigation localizes errors and provides recovery hints through a gold-blind auditor; the student preserves the valid prefix and regenerates the suffix. (d) Admitted trajectories update the training set, retaining valid historical targets; the next model trains from $M_{0}$ . (e) Fixed-set feedback guides question selection for the next round.

	Round-wise accuracy				Paired bootstrap
Benchmark	R0	R1	R2	R3	$\Delta$ (R3–R0)	95% CI
PMC-VQA	59.20	59.54 $\uparrow$	60.97 $\uparrow$	62.68 $\uparrow$	+3.48	[+1.9, +5.1]
VQA-RAD	68.10	68.47 $\uparrow$	68.06 $\downarrow$	70.40 $\uparrow$	+2.30	[+0.8, +3.8]
SLAKE	76.80	76.92 $\uparrow$	77.34 $\uparrow$	78.23 $\uparrow$	+1.43	[+0.2, +2.7]
PathVQA	48.80	52.38 $\uparrow$	51.27 $\downarrow$	53.34 $\uparrow$	+4.54	[+2.6, +6.5]
OmniMedVQA	82.20	80.81 $\downarrow$	81.58 $\uparrow$	85.41 $\uparrow$	+3.21	[+1.5, +4.9]

Table 3: Evolution and statistical uncertainty (Qwen3.5-9B). Left: accuracy (%) across rounds, each trained from Base (R0). Right: absolute gain R3

-

R0 and 95% bootstrap confidence intervals (CIs) (10,000 paired question resamples). The mean CI resamples each benchmark independently before averaging. All gain CIs exclude zero. Arrows compare consecutive rounds; MedXQA annotations give absolute changes. Bold/underline: best/second-best per row.

	35B				397B
Benchmark	R0	R1	R2	R3	R0	R1	R2
PMC-VQA	62.40	63.12 $\uparrow$	60.33 $\downarrow$	64.63 $\uparrow$	63.47	64.87 $\uparrow$	65.72 $\uparrow$
VQA-RAD	70.50	62.08 $\downarrow$	70.91 $\uparrow$	75.83 $\uparrow$	74.72	76.05 $\uparrow$	77.15 $\uparrow$
SLAKE	70.10	78.14 $\uparrow$	79.18 $\uparrow$	80.30 $\uparrow$	77.38	78.89 $\uparrow$	79.10 $\uparrow$
PathVQA	49.15	52.71 $\uparrow$	50.87 $\downarrow$	56.01 $\uparrow$	52.49	58.51 $\uparrow$	57.90 $\downarrow$
OmniMedVQA	83.20	85.39 $\uparrow$	85.22 $\downarrow$	83.97 $\downarrow$	85.10	85.15 $\uparrow$	85.05 $\downarrow$

Table 6: Scaling to larger models. Accuracy (%) for Qwen3.5-35B (left) and Qwen3.5-397B (right). Every round trains from Base (R0). Avg.: seven-benchmark mean. Arrows compare consecutive rounds; MedXQA annotations give absolute changes. Bold/underline: best/second-best within each model scale and benchmark.

Figure 5: Selection, hints, and feedback (Qwen3.5-9B). (a) Longest-8k versus random-8k questions: lower repair success can yield higher downstream accuracy. (b) Hint-count ablation uses the same longest-8k subset and run configuration as (a), with five continuation attempts;

k=1

reproduces its 66.78% single-hint baseline. Three hints perform best. (c,d) Training accuracy and seven-benchmark mean with training-error or fixed-set feedback. Feedback: 5,000 held-out questions.

Figure 6: Scale, uncertainty, and difficulty. (a) Mean accuracy after R3 for 9B/35B and R2 for 397B. (b) Repair acceptance with/without entropy localization; pointwise 95% Wilson intervals use 45,000 attempts per setting. (c) Error rates and step counts by min–max-normalized step-entropy bin; the diagonal is

y=100x

. (d) Accuracy on a modality-stratified 2,000-question subset of the feedback set; counts use the right axis. Heuristic is a training-free HHD-inspired hint-recovery baseline (Appendix 8.6 ).

		Agreement		Error family
Audit criterion	Pass (%)	$\kappa$	95% CI	V	R	D	K
Original trajectory erroneous	95.5	0.90	[0.83, 0.97]	96	96	94	96
Valid repair boundary	84.5	0.85	[0.66, 0.88]	88	84	82	84
No answer leakage in hint	97.2	0.94	[0.87, 1.00]	96	100	96	97
Image-grounded repaired reasoning	78.5	0.70	[0.58, 0.82]	76	80	76	82
Correct repaired answer	88.3	0.89	[0.81, 0.97]	88	89	88	88

Table 7: Human audit of repaired trajectories. Adjudicated binary pass rates (%) and pre-adjudication Cohen's

\kappa

. The 95% bootstrap CIs are for

\kappa

, resampling trajectories with the two independent ratings paired. V/R/D/K: visual, reasoning, differential-diagnosis, and knowledge errors. Repaired-output checks are shaded blue.

Resource	Size	Role
Training bank	81,293 questions	Source of original correct rollouts, accepted repairs, and retained valid targets.
Fixed feedback set	5,000 questions	Medical capability diagnosis and subsequent data selection; excluded from training.
Final evaluation	7 benchmarks	Independent downstream accuracy reporting.

Table 8: Dataset roles and sizes. Test coverage is counted in benchmarks.

Figure 7: Error-subtype evolution (Qwen3.5-9B). Bar heights normalize each subtype rate by R0; labels show original percentages. Arrows and multipliers compare R3 with R0. Family ratios divide summed R3 rates by summed R0 rates. Horizontal R1–R5 codes identify reasoning-error subtypes; legend R0–R3 labels identify training rounds.

Table 12

Subtype	R0	R1		R2		R3
V1 Missed key finding	3.12	2.95	$\downarrow$	2.36	$\downarrow$	2.12	$\downarrow$
V2 Spurious finding	2.87	2.82	$\downarrow$	2.28	$\downarrow$	2.05	$\downarrow$
V3 Location / laterality	2.65	2.68	$\uparrow$	2.20	$\downarrow$	1.98	$\downarrow$
V4 Sequence / phase	2.91	2.74	$\downarrow$	2.32	$\downarrow$	2.08	$\downarrow$
V5 Cross-image alignment	2.55	2.60	$\uparrow$	2.24	$\downarrow$	2.00	$\downarrow$
R1 Omitted evidence	2.88	2.55	$\downarrow$	2.16	$\downarrow$	1.95	$\downarrow$

Table 11: Error-subtype rates (%). Visual/reasoning subtypes are on the left; differential/knowledge subtypes on the right. R1–R5 are reasoning-error codes. Arrows compare consecutive rounds: blue

\downarrow

/red

\uparrow

/gray

=

indicate lower/higher/unchanged rates. Bold/underline: two lowest distinct rates per row; blue shading: R3.

Error family	R1			R2			R3
	Attempted	Successful	Success (%)	Attempted	Successful	Success (%)	Attempted	Successful	Success (%)
Visual misreading	12,189	9,826	80.61	10,245 $\downarrow$	6,789 $\downarrow$	66.27 $\downarrow$	10,478 $\uparrow$	7,094 $\uparrow$	67.70 $\uparrow$
Reasoning error	9,355	5,279	56.43	7,857 $\downarrow$	6,246 $\uparrow$	79.50 $\uparrow$	6,789 $\downarrow$	5,647 $\downarrow$	83.18 $\uparrow$
Differential error	4,064	2,995	73.70	6,742 $\uparrow$	4,673 $\uparrow$	69.31 $\downarrow$	4,786 $\downarrow$	3,291 $\downarrow$	68.76 $\downarrow$
Knowledge gap	1,625	1,467	90.28	987 $\downarrow$	550 $\downarrow$	55.72 $\downarrow$	1,567 $\uparrow$	765 $\uparrow$	48.82 $\downarrow$

Table 12: Repair outcomes by error family. Attempted and successful question counts, and repair success rates (%). Gray arrows track counts; blue/red arrows mark rising/falling success rates. Bold/underline: best/second-best rates per family; blue shading: R3.

Configuration	Value	Configuration	Value
Framework	ms-swift + Megatron-LM backend	add_non_thinking_prefix	true
Precision	BF16	loss_scale	ignore_empty_think
micro_batch_size	1	MAX_PIXELS	3145728
global_batch_size	64	dataset_num_proc	96
packing	true	dataloader_num_workers	8
attention_backend	flash	logging_steps	1

Table 13: MoE training configuration. Recorded settings for model training.

Medical Reasoning

Medical reasoning studies how AI systems use clinical evidence, patient context, and multimodal data to make and explain medical decisions.

See all 31 papers

Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning
Jun 30, 2026Junha Jung, Minbyul Jeong, Suhyeon Lim +5Process Reward Models Medical VQA
Korea University · AIGEN Sciences · Upstage AI +3
MedZERO: Self-Evolving Agents for Open-Ended Medical Reasoning Through Controlled Knowledge Accumulation
Oct 6, 2026Xilin Dang, Weilin Ruan, Xue Yang +4Self-Evolving Agents Medical LLMs
Department of Computer Science and Engineering, The Chinese University of Hong Kong · Department of Hematology, Zhongshan Hospital, Fudan University · School of Future Technology, South China University of Technology +1
Evo-PI: Aligning Medical Reasoning via Evolving Principle-Guided Supervision
Jun 30, 2026Xianda Zheng, Huan Gao, Meng-Fen Chiang +3LLM Alignment Medical VLMs
School of Computer Science, University of Auckland · Sapiens · National Yang Ming Chiao Tung University +3

Jun 30, 2026J/K move · Enter open · S save

Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning

Junha Jung, Minbyul Jeong, Suhyeon Lim, Sungwook Jung+4

Korea University · AIGEN Sciences · Upstage AI+3

Process Reward Models Medical VQA Multimodal Reasoning Medical VLMs RL for VLMs

Open paper AskPDF arXiv

Recent multimodal large language models have shown great promise in clinical image reasoning, but existing post-training pipelines remain predominantly outcome-centric, relying on final answer correctness or sequence-level preferences. This suffers from sparse credit assignment, making it difficult to optimize the reasoning process essential for clinical applications. Our analysis reveals that cascading errors from early-stage reasoning failures are a leading cause of incorrect predictions in medical visual question answering (VQA) benchmarks. Motivated by this, we propose Medical Reasoning-aware Policy Optimization (MRPO), an RL algorithm that incorporates step-wise process rewards. When the final answer is incorrect, MRPO assigns exponentially larger penalties to tokens in earlier invalid reasoning steps, breaking failure cascades without compromising successful paths. Across four multimodal LLM backbones, MRPO consistently outperforms standard GRPO and a recent RL baseline, and on Qwen3-VL-8B-Thinking even surpasses substantially larger medical MLLMs such as HuatuoGPT-Vision-34B by 4.59 points. Moreover, MRPO reduces early-stage reasoning failures from 58.6% to 13.4%, showing that targeted mitigation of cascading failures improves both reasoning quality and final answer accuracy. Our code is available at https://github.com/dmis-lab/MRPO

Abstract

Medical Reasoning

Explore similar work

Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning

MedZERO: Self-Evolving Agents for Open-Ended Medical Reasoning Through Controlled Knowledge Accumulation

Evo-PI: Aligning Medical Reasoning via Evolving Principle-Guided Supervision