MedEvo: Medical Reasoning Evolution through Knowledge Navigation
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.
Related
- Neural Structural Reasoner: A Brain-inspired Architecture for Reasoning over Structured Knowledge
- Agentic LLM reasoning in pediatric diabetes outcome prediction from multimodal medical records
- Seeing Through Uncertainty: Free-Energy-Inspired Real-Time Adaptation for Robust Visual Navigation
- CircuitATLAS: Agentic reasoning over a systems neuroscience knowledge graph for target discovery in circuitopathies
- The Cultural Evolution of Analytic and Holistic Thought: How Informational Ecologies shape Bayesian Reasoning
BibTeX
@misc{wang2026medevo,
title = {{MedEvo: Medical Reasoning Evolution through Knowledge Navigation}},
author = {Xiaoliang Wang and Zhuo Chen and Chuanyang Zheng and Quanbin Wang and Jiangwei Lao and Yingao Liu and Jiajia Liu and Jinming Cao and Honglin Lin and Zhihong Lu and Hao Jin and Jian Wang and Jinjie Gu and Lijun Wu},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.17690},
eprint = {2610.17690},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.17690},
}
RIS
TY - UNPB
TI - MedEvo: Medical Reasoning Evolution through Knowledge Navigation
AU - Xiaoliang Wang
AU - Zhuo Chen
AU - Chuanyang Zheng
AU - Quanbin Wang
AU - Jiangwei Lao
AU - Yingao Liu
AU - Jiajia Liu
AU - Jinming Cao
AU - Honglin Lin
AU - Zhihong Lu
AU - Hao Jin
AU - Jian Wang
AU - Jinjie Gu
AU - Lijun Wu
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2610.17690
UR - https://arxiv.org/abs/2610.17690
AB - 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.
ER -