Med-OPD: Improving Medical Vision-Language Models via Evidence-Aware On-Policy Distillation

2026-07-14

AI Summary

This work addresses the tendency of existing medical vision-language models to rely on linguistic priors rather than critical visual evidence for diagnosis. To mitigate this, the authors propose the Med-OPD framework, which introduces, for the first time, a Medical Evidence Advantage (MEA) signal. This signal quantifies evidence importance via counterfactual contrastive reasoning and dynamically adjusts intra-distillation weights at both token and trajectory levels to steer the student model toward visually diagnostic regions. By integrating evidence-degraded image modality contrast with multimodal modeling, the method substantially outperforms standard fine-tuning and conventional distillation across multiple OmniMedVQA subsets. It enhances the model’s reliance on key visual evidence and improves multimodal reasoning reliability in tasks involving CT and MRI imaging, disease diagnosis, and lesion grading.

Abstract

Medical Vision-Language Models (Med-VLMs) require reliable reasoning from fine-grained visual evidence, yet existing models can produce plausible clinical answers by relying on language priors or medical templates rather than truly attending to diagnosis-critical regions. On-Policy Distillation (OPD) offers dense token-level supervision on student-generated trajectories and provides a privacy-compatible means of capability transfer without requiring the redistribution of raw patient data. However, standard OPD uniformly distills all tokens, causing sparse evidence-dependent tokens to be diluted by abundant clinical narrative tokens. Inspired by the success of OPD in the large language model community, we propose \textbf{Med-OPD}, to our knowledge the first unified post-training framework that integrates on-policy distillation with medical evidence-aware supervision for Med-VLMs. We introduce \textbf{Medical Evidence Advantage} (MEA), a teacher-grounded counterfactual signal that uses an answer-aware hint to focus teacher scoring on evidence supporting the target diagnosis, and measures each token's dependence on medical visual evidence by comparing teacher likelihoods under the original and evidence-degraded imaging modalities. Based on MEA, Med-OPD redistributes the distillation signal at both the token and trajectory levels, emphasizing diagnosis-critical tokens and evidence-reliant rollouts. Experiments on OmniMedVQA subsets show that Med-OPD consistently outperforms SFT and standard OPD across CT, MRI, Disease Diagnosis, and Lesion Grading. These results demonstrate that evidence-aware distillation can better strengthen medical VLMs' reliance on key visual evidence and improve reliable multimodal medical reasoning. The source code and data is publicly available at: https://github.com/yunhang8658/MedOPD.git

Problem

• Medical Vision-Language Models
• visual evidence
• diagnosis-critical regions
• language priors
• reliable reasoning

Innovation

• Medical Vision-Language Models
• On-Policy Distillation
• Evidence-Aware Supervision
• Medical Evidence Advantage
• Multimodal Medical Reasoning