Medical reasoning agents built from synthetic knowledge trajectories

What it does

MedResearcher-R1 is an end-to-end framework for manufacturing training data for domain-specific reasoning models. It starts by building a knowledge graph and generating complex question-answer pairs with automated reasoning paths, then synthesizes multi-turn agent trajectories that include tool use, filters them for quality, and rewrites them into polished training examples. A final evaluation pipeline benchmarks the resulting model. The medical domain is the headline use case, but the architecture is framed as a general recipe for any deep-research specialty.

The interesting bit

Rather than scraping existing text, the system algorithmically extracts subgraphs from a curated knowledge graph to create questions, answers, and step-by-step “cheat sheets” that guide trajectory generation. A Masked Trajectory Guidance system then rewrites raw agent rollouts into cleaner training examples. It is essentially a synthetic curriculum designer where the structure of the domain knowledge directly shapes the reasoning patterns the model learns.

Key highlights

• End-to-end pipeline spanning KnowledgeGraphConstruction, TrajectoryGenerationPipeline, and EvaluationPipeline
• Five subgraph sampling strategies—mixed, augmented_chain, community_core_path, dual_core_bridge, and max_chain—for multi-hop question generation
• Automated quality filtering and LLM-powered trajectory rewriting with Masked Trajectory Guidance
• Ships with an open-source medical QA dataset and a D3.js web frontend for graph visualization
• Benchmarked on MedBrowseComp, GAIA, and XBench-DeepSearch

Caveats

• The built-in read tool hard-codes an OpenRouter dependency; using another provider requires editing tools/tool_visit.py
• Operating the full pipeline means managing separate configs and environment variables across three distinct component directories

Verdict

Worth exploring if you need a systematic way to generate synthetic reasoning trajectories from structured knowledge for fine-tuning domain agents. Skip it if you are looking for a drop-in medical chatbot—this is a data-production workshop, not a finished product.