| 考虑反思评估的医学大模型推理链方法 |
| Reasoning Chain Method for Medical Large Language Models Considering Reflective Evaluation |
| 摘要点击 32 全文点击 0 投稿时间：2025-12-21 修订日期：2026-06-14 |
| 查看/发表评论 下载PDF阅读器 |
| 中文关键词 医疗大模型；思维树；推理链增强；整体评估 |
| 英文关键词 Medical large language models; Tree of thought; Reasoning chain enhancement; Holistic evaluation |
| 基金项目 国家自然科学基金资助项目（72131006,72271082,72571086,72501088），安徽省自然科学基金杰青项目（2408085J041） |
| 投稿方向 信息系统工程 |
| 作者 | 单位 | 邮编 | | 樊怡平 | 合肥工业大学 | 230009 | | 顾东晓* | 合肥工业大学 | 230009 | | 苏凯翔 | 合肥工业大学 | | | 梁昌勇 | 合肥工业大学 | | | 王晓玉 | 安徽中医药大学第一附属医院 | |
|
| 中文摘要 |
| 现有医学大模型的思维链与思维树方法虽引入反思与验证机制, 但缺乏对长诊断推理链中各步骤间医学逻辑一致性的系统校验. 为此, 本研究提出一种结合思维树与推理链反思评估的医学诊断大模型推理链增强方法——RERCE-MDLLM. 本方法首先基于多阶段自适应思维树生成多样化的诊断推理链, 然后从医学知识一致性与关键证据权重两个方面对推理链进行综合评估以实现推理链增强, 并利用增强推理链进行直接偏好优化训练. 实验结果表明, 本方法优于通用指令模型、推理模型Marco-o1与医学专用模型Meditron3-8B. |
| 英文摘要 |
| Current medical large language models (LLMs) employ Chain-of-Thought and Tree-of-Thought reasoning strategies that incorporate reflection and verification mechanisms; however, they lack systematic validation of medical logical consistency across steps in long diagnostic reasoning chains. To address this limitation, this study proposes a reasoning chain enhancement method for medical diagnostic LLMs, termed RERCE-MDLLM, which integrates Tree-of-Thought generation with reasoning-chain reflection evaluation. The method first generates diverse diagnostic reasoning chains through a multi-stage adaptive Tree-of-Thought process. These reasoning chains are then comprehensively evaluated from the perspectives of medical knowledge consistency and key evidence weighting to enhance their overall quality,and the enhanced reasoning chains are subsequently used for Direct Preference Optimization training. The experimental results show that the proposed method outperforms general instruction-tuned models, the reasoning-oriented model Marco-o1, and the medical domain model Meditron3-8B. |
| 关闭 |
