Source-linked AI summary
MedResearcher-R1: Expert-Level Medical Deep Researcher via A Knowledge-Informed Trajectory Synthesis Framework
Ailing Yu, 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
TL;DR
General-purpose deep-research agents lack dense medical knowledge and specialized retrieval for complex queries. MedResearcher-R1 builds multi-hop medical training trajectories from rare-entity knowledge graphs and combines specialized tools with staged training; it reports state-of-the-art MedBrowseComp performance and robust general-benchmark results. The authors conclude that medical reasoning patterns can transfer to general tasks, while noting persistent knowledge-obsolescence and evidence-alignment problems in medical RAG systems.
Problem
General-purpose agents lack dense medical knowledge and specialized retrieval, while exploratory research requires connecting rare entities and synthesizing evidence across sources.
Source: §1 · p. 2Method
MedResearcher-R1 generates multi-hop trajectories from rare-entity knowledge graphs and trains with supervised fine-tuning, trajectory masking, and reinforcement learning.
Source: §1 · p. 3, §3.1.1 · p. 5, §7 · p. 11Results
MedResearcher-R1 achieves state-of-the-art MedBrowseComp pass@1 accuracy of 27.5% and robust performance on general agent benchmarks.
Source: §7 · p. 11Takeaways & Limitations
The authors report that dense medical knowledge structures and complex reasoning patterns transfer effectively to general tasks, suggesting specialization need not limit versatility.
Source: §1 · p. 3Medical RAG systems remain vulnerable to knowledge obsolescence and evidence misalignment; KBLaM’s MRR drops 18.4% after five update iterations.
Source: §6.2 · p. 11
Abstract
from arXiv · showhide
Recent developments in Large Language Model (LLM)-based agents have shown impressive capabilities spanning multiple domains, exemplified by deep research systems that demonstrate superior performance on complex information-seeking and synthesis tasks. While general-purpose deep research agents have shown impressive capabilities, they struggle significantly with medical domain challenges, as evidenced by leading proprietary systems achieving limited accuracy on complex medical benchmarks. The key limitations are: (1) the model lacks sufficient dense medical knowledge for clinical reasoning, and (2) the framework is constrained by the absence of specialized retrieval tools tailored for medical contexts. We present a medical deep research agent that addresses these challenges through two core innovations. First, we develop a novel data synthesis framework using medical knowledge graphs, extracting the longest chains from subgraphs around rare medical entities to generate complex multi-hop question-answer pairs. Second, we integrate a custom-built private medical retrieval engine alongside general-purpose tools, enabling accurate medical information synthesis. Our approach generates 2100+ diverse trajectories across 12 medical specialties, each averaging 4.2 tool interactions. Through a two-stage training paradigm combining supervised fine-tuning and online reinforcement learning with composite rewards, our MedResearcher-R1-32B model demonstrates exceptional performance, establishing new state-of-the-art results on medical benchmarks while maintaining competitive performance on general deep research tasks. Our work demonstrates that strategic domain-specific innovations in architecture, tool design, and training data construction can enable smaller open-source models to outperform much larger proprietary systems in specialized domains.
1 Introduction
General-purpose deep-research agents face gaps in dense medical knowledge and specialized retrieval for complex clinical queries. MedResearcher-R1 combines rare-entity training data, medical tools, and tailored learning, reporting strong medical benchmark results alongside competitive general-task performance.
Motivation
- General-purpose agents lack dense specialized medical knowledge and suitable retrieval tools; existing systems also struggle with exploratory research across rare entities and disparate evidence.OpenAI’s o3-deepresearch achieves 25.5% accuracy on complex medical queries requiring multi-hop reasoning.Source: §1 · p. 2
Data synthesis
- The data pipeline identifies rare, clinically significant entities and builds knowledge graphs around them to extract long reasoning chains for multi-hop questions.It analyzes more than 30 million PubMed abstracts and filters candidates with occurrence rates below 10^-6.Source: §1 · p. 2
Medical tools
- MedResearcher-R1 combines a private medical retrieval engine for authoritative sources with general-purpose tools to support evidence-chain validation.The medical sources include FDA databases, prescription data, clinical trial registries, and peer-reviewed publications.Source: §1 · p. 3
Training
- Knowledge-anchored training uses supervised fine-tuning on medical trajectories and Masked Trajectory Guidance to scaffold reasoning while preventing memorization.The authors report that initial supervised fine-tuning improves tool-use patterns and final performance.Source: §1 · p. 3
Results
- 27.5/50 on MedBrowseComp establishes state-of-the-art performance, while GAIA: 53.4 and xBench: 54 remain comparable to OpenAI o4-mini.The authors report that specialized medical training does not compromise general capabilities.Source: §1 · p. 3
Implications
- The authors report that dense medical knowledge structures and complex reasoning patterns transfer effectively to general tasks, suggesting specialization need not limit versatility.They frame tailored data, tools, and learning algorithms as necessary for medical deep research and release their code, datasets, and trained models.Source: §1 · p. 3
2 MedResearcher-R1: Medical Deep Research Agent Framework
The framework models medical research as sequential decision-making and equips an agent to combine general-purpose tools with specialized clinical retrieval and reasoning. Its policy selects tools as the agent iteratively gathers and incorporates evidence.
2.1 Problem Definition
- The agent models its evolving state through dialogue context, retrieved medical knowledge structured as a graph, and reasoning history, then selects actions with a learned policy.The state representation tracks explored knowledge paths and hypothesis evolution.Source: §2.1 · p. 3
2.2 Agent Architecture
- The agent follows iterative reason–act–observe cycles, formulating hypotheses, invoking tools, and incorporating structured evidence while pursuing multiple hypotheses.Complex multi-hop questions typically require 4–5 tool interactions.Source: §2.2 · p. 4
- The medical-specific suite combines authoritative-source retrieval with a Bayesian clinical reasoning engine that evaluates multiple diagnostic hypotheses.Retriever scores balance semantic relevance and clinical authority; diagnosis probabilities are updated using newly retrieved evidence.Source: §2.2 · pp. 4–5
- A learned policy selects between medical and general tools based on query complexity, prioritizing medical tools for rare diseases or complex compounds while retaining general tools for context.Selection features include entity rarity, required reasoning hops, and medical terminology.Source: §2.2 · pp. 4–5
3 KISA: Knowledge-Informed Trajectory Synthesis Approach
KISA synthesizes complex, multi-hop medical reasoning trajectories by centering training data on rare medical-entity connections and medical-specific retrieval tools.
3.1 Agentic Dataset Construction
- The dataset construction pipeline combines three components to generate complex medical queries that stress-test agent capabilities.The supplied passages specify the rare-entity graph construction, longest-path question generation, and difficulty calibration components.Source: §3.1 · p. 5
3.1.1 Entity-Centric Knowledge Graph Construction
- Rare entities below a 10^-6 frequency threshold seed knowledge graphs, while private medical retrieval discovers new entities for medically valid, challenging connections.The graph expansion also uses existing entity neighbors.Source: §3.1.1 · p. 5
- 12.3%: the enriched relation representation improves multi-hop reasoning accuracy compared with standard triplets.It encodes clinical context, temporal aspects, and spatial context.Source: §3.1.1 · p. 5
3.1.2 Multi-Hop Question Generation via Longest-Path Extraction
- Longest medically valid paths through rare-entity subgraphs become multi-hop questions requiring sequential tool calls to reconstruct the reasoning chain.Trajectories average 4.2 reasoning hops.Source: §3.1.2 · p. 5, Figure 3 · p. 6
3.1.3 Quality Control and Difficulty Calibration
- Questions are regenerated at greater complexity if either OpenAI-o3 deepresearch or GPT-4 scores above 50%, keeping the dataset challenging for current systems.The authors say this addresses the 25.5% performance ceiling previously observed in MedBrowseComp.Source: §3.1.3 · p. 6
3.2.1 Masked Trajectory Guidance (MTG)
- Masked Trajectory Guidance uses entity masking as a structural scaffold intended to support reasoning without shortcut learning.Source: Figure 4 · p. 6
3.2.2 Hybrid Strategy for Tool Diversity
- 78% begin with private medical retriever for rare entities, 42% demonstrate adaptive switching between general and medical tools, and 34% include explicit correction using alternative tools.The hybrid strategy sets α = 0.7 to balance structured learning with exploration.Source: §3.2.2 · p. 7
4 Large-scale Agent Training
Training begins with supervised fine-tuning on synthetic agent dialogues, then uses reinforcement learning to refine trajectories against composite rewards. The two stages incorporate robustness and multi-task augmentations, curriculum learning, and penalties for unnecessary tool use.
4.1 Cold Start with Supervised Fine-Tuning
- SFT maximizes the likelihood of the correct next-action trajectory given the input context and prior agent history.Examples pair an input context with an ideal sequence of thoughts, tool calls, and other actions.Source: §4.1 · p. 7
- Training augments trajectories with 5% corrupted tool outputs, intermediate-thought supervision, and multi-task sampling across medical domains.The listed domains include diagnosis, treatment, guidelines, and rare diseases.Source: §4.1 · p. 7
4.2 Reinforcement Learning
- After SFT, GRPO optimizes trajectories using task accuracy, GPT-4 expert preference, and an efficiency penalty for unnecessary or irrelevant tool use.The reward weights are 1.0, 0.2, and 0.1, respectively.Source: §4.2 · p. 7
- RL increases task complexity progressively using a pass-rate-monitored curriculum, and removes KL regularization because it may hinder multi-stage performance gains.The group-level baseline is the average reward among responses in the same batch.Source: §4.2 · p. 8
5 Experiments
The experiments evaluate medical evidence retrieval and synthesis alongside general agent capabilities. MedResearcher-R1 achieves a new MedBrowseComp state of the art, while analyses associate iterative search and verification with stronger multi-hop performance.
5.1 Benchmarks
- MedBrowseComp tests open-ended clinical questions requiring multi-step reasoning, strategic information gathering, and synthesis of medical evidence from multiple web sources.GAIA evaluates real-world multimodal assistant tasks, while XBench-DeepSearch assesses tool use across open-domain tasks.Source: §5.1 · p. 8
5.2 Main Results
- 27.5/50 pass@1 on MedBrowseComp establishes a new state of the art, outperforming previous best agents and the Qwen2.5-32B baseline.SFT delivers gains, and subsequent reinforcement learning further improves decision quality and tool orchestration efficiency.Source: §5.2 · p. 8, Table 1 · p. 8
- On GAIA and XBench-DeepSearch, MedResearcher-R1 earns competitive helpfulness scores, indicating generalization beyond medical tasks.Table 2 compares performance on these two general agent benchmarks.Source: §5.2 · p. 8, Table 2 · p. 9
5.3 Qualitative Analysis
- Training instances following iterative search-verification-synthesis show 34.2% higher success rates on complex multi-hop tasks than single-verification approaches.The described strategy searches broadly, checks consistency across authoritative medical databases, follows up on ambiguities, then synthesizes validated findings.Source: §5.3 · pp. 9–10
6 Related Work
Related work covers general deep-research agents, medical retrieval-augmented systems, and medical multi-role agents. The discussion highlights their capabilities as well as remaining limits in healthcare evidence handling and reasoning depth.
6.1 General Deep Research Methods
- General deep-research methods use multi-agent planning or reward-trained agents, but prompt-led planning can propagate errors and general web agents lack healthcare-specific safeguards.The cited healthcare gaps include de-identification, clinical evidence grading, and medication compliance audits.Source: §6.1 · p. 10
6.2 Medical RAG Systems
- Medical RAG systems use curated or live evidence sources, yet manual updates can delay retrieval relevance and KBLaM embedding misalignment lowers MRR by 18.4% after five iterations.DeepRare [Zhao et al., 2025a] is cited for manual update orchestration; the MRR decline is attributed to Wang et al. [Wang et al., 2025].Source: §6.2 · pp. 10–11
6.3 Medical Multi-Role Systems
- Medical multi-role systems integrate dynamic knowledge and unified pipelines, yet AgentClinic drops 27.3% at seven reasoning steps and MAI-DxO degrades 34% in deterioration scenarios.The discussion also identifies deficits in causal treatment planning.Source: §6.3 · p. 11
7 Conclusion
The conclusion presents KISA as a way to construct challenging multi-hop medical research examples from rare entities and knowledge-graph reasoning chains. MedResearcher-R1 combines this data with specialized tools and staged training, achieving a reported MedBrowseComp state of the art.
7 Conclusion
- KISA generates multi-hop medical question-answer pairs and reasoning trajectories from rare-entity mining and knowledge-graph reasoning chains.The approach aims to expose agents to intricate, compositional problems characteristic of real-world medical research.Source: §7 · p. 11
- MedResearcher-R1 combines the dataset with supervised fine-tuning, trajectory masking, reinforcement learning, and specialized medical tools.Source: §7 · p. 11
- 27.5% pass@1 accuracy on MedBrowseComp is reported as state of the art, alongside robust performance on general agent benchmarks.The conclusion describes the model as capable of complex medical questions requiring systematic exploration and nuanced evidence synthesis.Source: §7 · p. 11
8 Future Work
Future work spans multimodal clinical inputs, clinician feedback, safety and reliability, and stronger multi-hop benchmarks; code, datasets, and trained models will be released to support collaboration and evaluation.
Multi-modal Tool Integration
- Multimodal integration would add radiology, pathology, genomic, and electronic-health-record tools to support synthesis across data types and better align with clinical workflows.Source: §8 · p. 12
Human-Expert Collaboration
- Human-in-the-loop feedback from medical professionals is proposed to improve reasoning quality, tool use, and the clinical relevance of agent outputs.Source: §8 · p. 12
Safety and Reliability
- Open deployment motivates systematic study of hallucination detection, uncertainty estimation, and fail-safe mechanisms for high-stakes medical scenarios.Source: §8 · p. 12
Advanced Medical Reasoning Benchmarks
- A comprehensive benchmark is proposed for multi-hop reasoning across pharmacology, diagnostics, epidemiology, genetics, surgical planning, and therapy.It would set a higher standard for evaluating tool orchestration and evidence synthesis in challenging scenarios.Source: §8 · p. 12
Future Work
- Releasing the codebase, datasets, and trained models is intended to foster collaborative progress and rigorous evaluation toward trustworthy healthcare research agents.Source: §8 · p. 12