cs.CV提交于 2026-10-27已译

GRASP：面向医学视觉问答、无需推理标注的组自适应混合 RL-SFT 后训练

GRASP: Group-Adaptive Hybrid RL-SFT Post-Training for Medical VQA without Reasoning Annotations

Wanqi Dong · Wanyue Dong · Jingze Ge · Mehul Motani

中文摘要

通用视觉语言模型(VLMs)适配医学视觉问答(VQA)，可辅助临床医生读图与诊断，但标注数据稀缺且仅含简短答案、没有推理标注。监督微调(SFT)会拟合这些标签，过拟合并放大偏差，削弱域外(OOD)泛化；强化学习(RL)更能保持泛化，却从不以标准答案作为训练目标。GRPO 类算法仅从一组采样回答中正确与错误回答的对比学习，若全对或全错便无正确性信号。现有混合方案要么分阶段执行或按问题切换 SFT 与 RL，要么依赖专家或更强模型提供推理轨迹，而医学 VQA 缺乏这些数据。GRASP 在每次更新中结合 RL 与 SFT，仅用简短答案：RL 探索模型自身推理，随后 SFT 用标准答案提供 RL 缺失的直接监督，并依据模型作答表现自适应平衡二者。通过 LoRA 仅调制约 1% 参数，GRASP 在 Qwen3-VL-8B、InternVL3-8B 和 Gemma 3-4B 的域内及 OOD 开放式问题上均优于所有对比训练目标。在覆盖 8 种成像模态的 OOD benchmark 上，经 GRASP 训练的 InternVL3-8B 与 Qwen3-VL-8B 开放式准确率超过全部 10 个通用及医学 VLM，其中包括训练数据多达 2,600 倍或可训练参数超过 250 倍的模型。项目同时发布含 12.9K 条简短答案问题的 MedTerse 数据集及代码。

关键要点

01SFT 易过拟合短答案标签并放大偏差，RL 也缺少标准答案的直接监督。
02GRASP 每次更新结合 RL 探索推理与 SFT 答案监督，并按问题自适应平衡。
03仅用 LoRA 调整约 1% 参数，即在多个医学 VQA 模型上取得更优结果。
04在覆盖 8 种成像模态的 OOD benchmark 上，开放式准确率超过全部 10 个对比 VLM。
05发布含 12.9K 条简短答案问题的 MedTerse 数据集与开源代码。

解读

尚无解读。

原始英文摘要

arXiv:2610.34561v1 Announce Type: new Abstract: Adapting general-purpose vision-language models (VLMs) to medical visual question answering (VQA) is valuable for helping clinicians read images and make diagnoses, but annotated data are scarce and provide only short answers, without reasoning annotations. Supervised fine-tuning (SFT) fits these labels but overfits to them and amplifies their biases, eroding out-of-domain (OOD) generalization. Reinforcement learning (RL) better preserves generalization but never uses the gold answer as a training target: GRPO-style algorithms learn from the contrast between correct and incorrect answers within a group of sampled responses, so all-correct or all-wrong groups give no correctness signal. Existing SFT-RL hybrids either apply SFT and RL in separate stages or switch between them per question, or need reasoning traces from experts or stronger models, which medical VQA data lack. We propose GRASP, which combines RL and SFT in every update using only short answers: RL lets the model explore its own reasoning, SFT teaches the gold answer after that reasoning to supply the direct supervision RL lacks, and their balance adapts to how well the model answers each question. Tuning only about 1% of the parameters with LoRA, GRASP outperforms all compared training objectives both in domain and on OOD open-ended questions across Qwen3-VL-8B, InternVL3-8B and Gemma 3-4B. On an OOD benchmark spanning eight imaging modalities, GRASP-trained InternVL3-8B and Qwen3-VL-8B beat all 10 compared general and medical VLMs in open-ended accuracy, including models trained on up to 2,600 times more data or with over 250 times more trainable parameters. We release MedTerse, a dataset of 12.9K short-answer questions, and our code at https://github.com/sea-comet/GRASP-MedVQA.