📄 中文摘要
大型语言模型（LLMs）在医疗问答中展现出较强的推理能力，但其产生幻觉和过时知识的倾向在医疗领域中带来了重大风险。虽然增强检索生成（RAG）方法能够缓解这些问题，但现有方法依赖于嘈杂的标记级信号，缺乏复杂推理所需的多轮精炼。MA-RAG（多轮智能 RAG）框架通过在智能精炼循环中迭代演变外部证据和内部推理历史，促进复杂医疗推理的测试时扩展。在每一轮中，代理将候选响应之间的语义冲突转化为可操作的查询，从而提升推理的准确性和可靠性。
从冲突到共识：通过多轮智能 RAG 提升医疗推理能力
📄 English Summary
From Conflict to Consensus: Boosting Medical Reasoning via Multi-Round Agentic RAG
Large Language Models (LLMs) demonstrate significant reasoning capabilities in medical question-answering; however, their propensity to generate hallucinations and outdated knowledge poses critical risks in the healthcare domain. While Retrieval-Augmented Generation (RAG) techniques help mitigate these issues, existing approaches rely on noisy token-level signals and lack the multi-round refinement necessary for complex reasoning tasks. The proposed MA-RAG (Multi-Round Agentic RAG) framework facilitates test-time scaling for intricate medical reasoning by iteratively evolving both external evidence and internal reasoning history within an agentic refinement loop. In each round, the agent transforms semantic conflicts among candidate responses into actionable queries, thereby enhancing the accuracy and reliability of the reasoning process.