摘要
针对大语言模型在医疗应用中缺乏高质量推理数据的问题，本文提出 MedSSR 框架。该方法利用罕见病知识合成可控分布的推理问题，并通过策略模型生成高质量伪标签，构建了从自监督到监督的两阶段强化学习范式。实验表明，该方法在不依赖昂贵思维链蒸馏的情况下，显著提升了模型在十个医疗基准测试中的表现，尤其在罕见病任务上增益达 5.93%。
AI 推荐理由
论文核心聚焦于通过知识增强数据合成与半监督强化学习提升医疗推理能力。
研究机构
College of Computer Science and Artificial Intelligence, Fudan University
CMIC, Shanghai Jiao Tong University
School of Artificial Intelligence, Shanghai Jiao Tong University
Department of Radiology, Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine
Shanghai AI Laboratory
Institute of Artificial Intelligence for Medicine, Shanghai Jiao Tong University School of Medicine
论文信息