When Does RL Help Medical VLMs? Disentangling Vision, SFT, and RL Gains
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2026
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866908859222720512 |
|---|---|
| author | Jeddi, Ahmadreza Shaban, Kimia Baghbanzadeh, Negin Sharan, Natasha Moturu, Abhishek Dolatabadi, Elham Taati, Babak |
| author_facet | Jeddi, Ahmadreza Shaban, Kimia Baghbanzadeh, Negin Sharan, Natasha Moturu, Abhishek Dolatabadi, Elham Taati, Babak |
| contents | Reinforcement learning (RL) is increasingly used to post-train medical Vision-Language Models (VLMs), yet it remains unclear whether RL improves medical visual reasoning or mainly sharpens behaviors already induced by supervised fine-tuning (SFT). We present a controlled study that disentangles these effects along three axes: vision, SFT, and RL. Using MedMNIST as a multi-modality testbed, we probe visual perception by benchmarking VLM vision towers against vision-only baselines, quantify reasoning support and sampling efficiency via Accuracy@1 versus Pass@K, and evaluate when RL closes the support gap and how gains transfer across modalities. We find that RL is most effective when the model already has non-trivial support (high Pass@K): it primarily sharpens the output distribution, improving Acc@1 and sampling efficiency, while SFT expands support and makes RL effective. Based on these findings, we propose a boundary-aware recipe and instantiate it by RL post-training an OctoMed-initialized model on a small, balanced subset of PMC multiple-choice VQA, achieving strong average performance across six medical VQA benchmarks. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_01301 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | When Does RL Help Medical VLMs? Disentangling Vision, SFT, and RL Gains Jeddi, Ahmadreza Shaban, Kimia Baghbanzadeh, Negin Sharan, Natasha Moturu, Abhishek Dolatabadi, Elham Taati, Babak Computer Vision and Pattern Recognition Reinforcement learning (RL) is increasingly used to post-train medical Vision-Language Models (VLMs), yet it remains unclear whether RL improves medical visual reasoning or mainly sharpens behaviors already induced by supervised fine-tuning (SFT). We present a controlled study that disentangles these effects along three axes: vision, SFT, and RL. Using MedMNIST as a multi-modality testbed, we probe visual perception by benchmarking VLM vision towers against vision-only baselines, quantify reasoning support and sampling efficiency via Accuracy@1 versus Pass@K, and evaluate when RL closes the support gap and how gains transfer across modalities. We find that RL is most effective when the model already has non-trivial support (high Pass@K): it primarily sharpens the output distribution, improving Acc@1 and sampling efficiency, while SFT expands support and makes RL effective. Based on these findings, we propose a boundary-aware recipe and instantiate it by RL post-training an OctoMed-initialized model on a small, balanced subset of PMC multiple-choice VQA, achieving strong average performance across six medical VQA benchmarks. |
| title | When Does RL Help Medical VLMs? Disentangling Vision, SFT, and RL Gains |
| topic | Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2603.01301 |