MedEvo: Medical Reasoning Evolution through Knowledge Navigation

Xiaoliang Wang; Zhuo Chen; Chuanyang Zheng; Quanbin Wang; Jiangwei Lao; Yingao Liu; Jiajia Liu; Jinming Cao; Honglin Lin; Zhihong Lu; Hao Jin; Jian Wang; Lijun Wu

doi:10.48550/arXiv.2610.17690

MedEvo: Medical Reasoning Evolution
through Knowledge Navigation

* Equal contribution† Corresponding author

Advancing medical reasoning through knowledge-guided self-improvement.

Paper ↗GitHub ↗

KNOWLEDGE NAVIGATION

Case 1/4 · Schematic

Medical reasoning repair and supervision evolution

V1 · A missed finding calls for another look at the relevant image region.

Four schematic cases demonstrate visual reinspection, evidence reconciliation, discriminative comparison, and a rule-category cue. Each continues through screened supervision, SFT from the same base, and feedback-guided selection across four fixed medical tag axes.

Four fixed medical tag axes guide selection. Repair uses observable tags; feedback questions stay out of training.

What the experiments reveal

Data selection and feedback

Harder to repair.
More useful to learn.

Qwen3.5-9B · one round
	Longest 8k	Random 8k
Repair success	55.70%	67.50%
After training	66.78%	64.35%

Mean benchmark accuracy. Length and difficulty are not separated.

Independent feedback
changes the direction.

Seven-benchmark mean · Qwen3.5-9B

Training-error feedback65.57 → 64.01%

Fixed-set feedback65.57 → 69.12%

Base → round 3. Better training fit need not transfer.

See the evidence in Figure 5

The paper's selection, hint-count, and independent-feedback comparisons, including longest-8k accuracy of 66.78% and fixed-set feedback accuracy of 69.12%. — Selection, hints, and feedback.

Medical multimodal benchmarks

Model comparison

Table 2 in the paper ↗

Medical model Trajectory repairScroll horizontally for all benchmarks →

Accuracy (%) · MedEvo after round 3
Model	PMC- VQA	VQA- RAD	SLAKE	PathVQA	OmniMed VQA	MedXQA	MMMU- Med	Avg.
Proprietary models
GPT-5.6-sol	60.43	70.29	81.34	54.84	76.67	77.50	86.24	72.47
Gemini-3.8-Flash	64.52	74.72	81.34	58.58	80.76	86.30	89.10	76.47
Open-source models (<10B)
LLaVA-Med-7B	30.50	53.70	48.00	38.80	44.30	20.30	29.30	37.84
HuatuoGPT-V-7B	53.30	67.00	67.80	48.00	74.20	21.60	47.30	54.17
BiMediX2-8B	43.50	49.20	57.70	37.00	63.30	21.80	39.80	44.61
InternVL2.5-8B	51.30	59.40	69.00	42.10	81.30	21.70	53.50	54.04
InternVL3-8B	53.80	65.40	72.80	48.60	79.10	22.40	59.20	57.33
LINGSHU-7B	56.30	67.90	83.10	61.90	82.90	26.70	54.00	61.83
Qwen3.5-9B (Base)	59.20	68.10	76.80	48.80	82.20	47.40	76.50	65.57
Baseline 1 (HHD)	58.93	67.65	71.45	50.12	82.35	48.75	76.90	65.16
Baseline 2 (TrajFusion)	59.32	69.12	77.23	49.23	83.46	49.56	77.20	66.45
Baseline 3 (Teacher SFT)	61.23	70.15	71.60	53.20	81.90	52.30	78.14	66.93
MedEvo (9B)	62.68	70.40	78.23	53.34	85.41	54.50	79.28	69.12
Open-source models (>10B)
HealthGPT-14B	56.40	65.00	66.10	56.70	75.20	24.70	49.60	56.24
HuatuoGPT-V-34B	56.60	61.40	69.50	44.40	74.00	22.10	51.80	54.26
InternVL3-14B	54.10	66.30	72.80	48.00	78.90	23.10	63.10	58.04
InternVL2.5-38B	57.20	61.40	70.30	46.90	79.90	24.40	61.60	57.39
InternVL3-38B	56.60	65.40	72.70	51.00	79.80	25.20	65.20	59.41
LINGSHU-32B	57.90	76.50	89.20	65.90	83.40	30.90	62.30	66.59
Qwen3.5-35B (Base)	62.40	70.50	70.10	49.15	83.20	60.00	80.00	67.91
MedEvo (35B)	64.63	75.83	80.30	56.01	83.97	66.50	84.02	73.04

Bold / underline: best / second-best within each source group, with shared ranks for ties. Avg.: reported unweighted seven-benchmark mean. MedXQA: MedXpertQA-MM; MMMU-Med: medical subset of MMMU.

SFT only

Three model scales

Seven-benchmark mean accuracy (%)

Qwen3.5-9B65.57 → 69.123 rounds · +3.55 pp

Qwen3.5-35B67.91 → 73.043 rounds · +5.13 pp

Qwen3.5-397B71.83 → 74.952 rounds · +3.12 pp

pp: percentage points.

Paper ↗GitHub ↗Project page ↗Related: ECG-RSI ↗

Research code and training data are not released.

Full method diagram

Full MedEvo method diagram from the paper. — Only input-observable tags reach the auditor. Feedback questions guide selection and never enter training.

Cite MedEvo

@misc{wang2026medevo,
  title  = {{MedEvo}: Medical Reasoning Evolution through Knowledge Navigation},
  author = {Wang, Xiaoliang and Chen, Zhuo and Zheng, Chuanyang and
            Wang, Quanbin and Lao, Jiangwei and Liu, Yingao and Liu, Jiajia and
            Cao, Jinming and Lin, Honglin and Lu, Zhihong and Jin, Hao and
            Wang, Jian and Wu, Lijun},
  year   = {2026},
  eprint = {2610.17690},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  doi = {10.48550/arXiv.2610.17690},
  url = {https://arxiv.org/abs/2610.17690}
}