DEEPMED: Building a Medical DeepResearch Agent via Multi-hop Med-Search Data and Turn-Controlled Agentic Training & Inference

TL;DR

DeepMed addresses hallucination and knowledge-forgetting in medical reasoning by grounding outputs in open-web medical evidence using a DeepResearch framework tailored for medicine. It combines a web-sourced multi-hop med-search data synthesis with two training stages—Agentic SFT and Agentic RL—plus a difficulty-aware turn-penalty and an Over-Evidence Monitor to curb over-searching. Across seven medical benchmarks, DeepMed achieves substantial gains over its base model and outperforms larger medical reasoning and DR models, with notable improvements on challenging datasets and strong evidence-grounding behavior. Analyses show that tool-augmented search–reflection–verification fosters self-correction and enhances the completeness and credibility of medical reasoning chains, suggesting practical potential for medically informed AI systems while highlighting computational and validation limitations. The work contributes a concrete blueprint for adapting DeepResearch to high-stakes domains and points to future directions in scalable data synthesis, efficient tool use, and clinical deployment.

Abstract

Medical reasoning models remain constrained by parametric knowledge and are thus susceptible to forgetting and hallucinations. DeepResearch (DR) models ground outputs in verifiable evidence from tools and perform strongly in general domains, but their direct transfer to medical field yields relatively limited gains. We attribute this to two gaps: task characteristic and tool-use scaling. Medical questions require evidence interpretation in a knowledge-intensive clinical context; while general DR models can retrieve information, they often lack clinical-context reasoning and thus "find it but fail to use it," leaving performance limited by medical abilities. Moreover, in medical scenarios, blindly scaling tool-call can inject noisy context, derailing sensitive medical reasoning and prompting repetitive evidence-seeking along incorrect paths. Therefore, we propose DeepMed. For data, we deploy a multi-hop med-search QA synthesis method supporting the model to apply the DR paradigm in medical contexts. For training, we introduce a difficulty-aware turn-penalty to suppress excessive tool-call growth. For inference, we bring a monitor to help validate hypotheses within a controlled number of steps and avoid context rot. Overall, on seven medical benchmarks, DeepMed improves its base model by 9.79\% on average and outperforms larger medical reasoning and DR models.

Figures (14)

• Figure 1: General search tasks typically provide definitive answers online, but medical tasks are different; while online information can assist, the final answer relies on medical reasoning (up). DeepMed shows significant improvements over its base reasoning model on medical benchmarks, and outperforms larger medical reasoning model and general DR models (down).
• Figure 2: The multi-hop Med-Search data synthesis workflow for Agentic SFT of DeepMed. The top panel illustrates how we synthesize multi-hop medical chains that emphasize logical relations by leveraging multi-source web evidence. The bottom panel shows how these chains are transformed into Med-Search QA.
• Figure 3: A full rollout and its reward evaluation are shown. Only segments inside the solid boxes contribute to the loss. Incorrect format or answer receives zero reward, and even correct rollouts are penalized for excessive rounds.
• Figure 4: The workflow of 'Over-Evidence' monitor.
• Figure 5: Answer outcome breakdown: most correct predictions stem from tool use and self-correction, while errors caused by later tool use or reasoning are rare.