MedResearcher-R1：基于知识驱动轨迹合成框架的专家级医学深度研究智能体
MedResearcher-R1: Expert-Level Medical Deep Researcher via A Knowledge-Informed Trajectory Synthesis Framework
MedResearcher-R1-32B 在 MedBrowseComp 上取得 27.5/50，超过 o3-deepresearch 的 25.5/50。该医学深度研究智能体采用知识驱动轨迹合成（KISA）方法构建医学知识图谱，生成覆盖 12 个医学专科的 2100 余条轨迹，并集成医学检索引擎与通用工具。模型经监督微调和强化学习训练，在 GAIA 和 xBench 上同样表现良好。
Authors:Ailing Yu, HaoYang Hong, Lan Yao, Jingnan Liu, Zhe Chen, Jiajun Yin, Yuan Wang, Xinhao Liao, Zhiling Ye, Ji Li, Yun Yue, Hansong Xiao, Hualei Zhou, Chunxiao Guo, Peng Wei, Junwei Liu, Jinjie Gu
Abstract:Large Language Model (LLM)-based agents have recently demonstrated remarkable capabilities in deep research tasks, yet their application to the medical domain remains challenging. The leading proprietary systems achieve only modest results on complex medical benchmarks. This exposes two key bottlenecks: (i) limited clinical-reasoning capability in the underlying model, and (ii) unreliable access to authoritative medical evidence when constrained to open-web retrieval. To address these limitations and equip agents with robust clinical reasoning capabilities, we present MedResearcher-R1, a medical deep research agent featuring two core innovations. First, we propose Knowledge-Informed Trajectory Synthesis (KISA), a novel approach that builds medical knowledge graphs to construct complex multi-hop question-answer pairs centered on rare medical entities, overcoming the scarcity of high-quality domain-specific training data. Second, we integrate a medical retrieval engine alongside general-purpose tools, enabling precise and reliable synthesis of medical evidence. Through this methodology, we yield over 2,100 diverse trajectories spanning 12 medical specialties. Trained with supervised fine-tuning and reinforcement learning, our MedResearcher-R1-32B achieves state-of-the-art performance on MedBrowseComp (27.5/50 vs. 25.5/50 for o3-deepresearch) while demonstrating strong general performance on GAIA and xBench benchmarks. To the best of our knowledge, we present the first high-quality, tool-augmented medical dataset paired with a domain-specialized deep-research agent, demonstrating that smaller open-source models can surpass substantially larger proprietary systems on specialized medical tasks.
| Comments: | 14 pages, 4 figures |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2508.14880 [cs.CL] |
| (or arXiv:2508.14880v4 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2508.14880 arXiv-issued DOI via DataCite |
Submission history
From: Ailing Yu [view email]
[v1]
Wed, 20 Aug 2025 17:51:20 UTC (5,716 KB)
[v2]
Thu, 21 Aug 2025 18:29:42 UTC (5,905 KB)
[v3]
Mon, 1 Sep 2025 15:33:47 UTC (5,905 KB)
[v4]
Thu, 29 Oct 2026 15:13:07 UTC (2,338 KB)
来源：arXiv cs.CL · arxiv.org