AQ-MedAI/MedResearcher-R1-32B

MedResearcher-R1: Expert-Level Medical Deep Researcher via A Knowledge-Informed Trajectory Synthesis Framework

Author List

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, Jinjie Gu

Abstract

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—the MedBrowseComp benchmark reveals even GPT-o3 deep research, the leading proprietary deep research system, achieves only 25.5% accuracy on complex medical queries. The key limitations are: (1) insufficient dense medical knowledge for clinical reasoning, and (2) lack of medical-specific retrieval tools. 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 longest chains from subgraphs around rare medical entities to generate complex multi-hop QA pairs. Second, we integrate a custom-built private medical retrieval engine alongside general-purpose tools, enabling accurate medical information synthesis. Our approach generates 2,100 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 open-source 32B model achieves competitive performance on general benchmarks (GAIA: 53.4, xBench: 54), comparable to GPT-4o-mini, while outperforming significantly larger proprietary models. More importantly, we establish new state-of-the-art on MedBrowseComp with 27.5% accuracy, surpassing leading closed-source deep research systems including O3 deepresearch, substantially advancing medical deep research capabilities. 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. Code and datasets will be released to facilitate further research.

MedResearcher-R1 is a comprehensive training data generation and synthesis framework that tackles the challenge of domain-specific AI reasoning through knowledge-informed trajectory synthesis. Our framework provides an end-to-end solution for generating high-quality training data, consisting of three integrated components:

🧠 Knowledge Graph Construction: Our core innovation - an intelligent knowledge graph construction and QA synthesis system that transforms domain knowledge into high-quality question-answer pairs with automated reasoning path generation. This module serves as the foundation for creating domain-specific training data.

🔄 Trajectory Generation Pipeline: End-to-end trajectory synthesis and optimization system that converts QA pairs into multi-turn reasoning trajectories with tool interactions and quality filtering for model training.

📊 Evaluation Pipeline: Comprehensive model evaluation and validation framework for assessing reasoning performance across multiple benchmarks and validating the quality of synthesized training data.

These three components form a complete training data production pipeline from knowledge extraction to model training data generation and evaluation, enabling the creation of specialized reasoning models for domain-specific applications.

Features

• Knowledge Graph Construction

  • Interface Support: Interactive web visualization with D3.js force-directed graphs
  • Advanced Sampling Algorithms: 5 sophisticated strategies (mixed, augmented_chain, community_core_path, dual_core_bridge, max_chain) for complex subgraph extraction
  • Unified QA Generation: Deep concept obfuscation with quantitative reasoning and multi-paradigm question synthesis
  • Reasoning Path Generation: Automated cheat_sheet creation with detailed step-by-step reasoning guidance for complex multi-hop questions
  • Batch Processing System: Concurrent QA generation with intelligent QPS control, progress monitoring, and resume capability

• Trajectory Generation Pipeline

  • Agent Framework: Multi-turn reasoning with tool integration and concurrent task processing
  • Advanced Quality Filtering: Token-based validation, tool call/response matching, and automated error detection
  • Intelligent Rewriting System: LLM-powered trajectory optimization with Masked Trajectory Guidance (MTG)

• Evaluation Pipeline

  • Interactive Question Reasoning: Single question mode with detailed step-by-step process visualization
  • Batch Dataset Evaluation: Multi-worker parallel processing with configurable rollouts and timeout controls

Performance Highlights

Using our knowledge-informed trajectory synthesis framework, we developed MedResearcher-R1, a specialized reasoning model that demonstrates exceptional performance across multiple challenging benchmarks including MedBrowseComp, GAIA, and XBench-DeepSearch.

Open-Sourced Dataset

We have open-sourced a high-quality QA dataset constructed through our KnowledgeGraphConstruction module. The dataset is available at TrajectoryGenerationPipeline/qa_data/open_data.jsonl and contains:

• Complex reasoning question-answer pairs Multi-hop qa-pairs generated using our graph method
• Detailed step-by-step reasoning paths for each question, providing comprehensive problem-solving guidance

Quick start: Run Model for Evaluation

You can run a server for the model via sglang or vllm for evaluation, as described in the GitHub repository's Quick start section.

First, install sglang (e.g., pip install sglang[all]):

pip install sglang[all]
CUDA_VISIBLE_DEVICES=0,1 python -m sglang.launch_server --model-path /path/to/your/model --port 6001 --host 0.0.0.0 --mem-fraction-static 0.95 --tp-size 2

Then, you can evaluate model performance using the Evaluation Pipeline as detailed in the GitHub repo:

cd ../EvaluationPipeline
# Run single question evaluation
python eval_cli.py --mode interactive

# Run batch dataset evaluation
python eval_cli.py --mode batch --dataset sample --workers 20

✍️ Citation

@article{ant2025medresearcher,
  title={MedReseacher-R1: Expert-Level Medical Deep Researcher via A Knowledge-Informed Trajectory Synthesis Framework},
  author={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, Jinjie Gu},
  journal={arXiv preprint},
  url={https://arxiv.org/abs/2508.14880},
  year={2025}
}

📜 License

MedReseacher-R1 is licensed under the MIT license.

---