HealthRecordCommunity
FHIR🌏 Germany🏛 IDM gGmbH, University Medical Center Hamburg-EppendorfEnriched

Reinforcement Learning for Tool-Calling Agents in Fast Healthcare Interoperability Resources (FHIR)

FHIRデータにおけるツール呼び出しエージェントのための強化学習

October 13, 2026IDM gGmbH, University Medical Center Hamburg-Eppendorf

Summary

Researchers addressed the challenge of multi-step reasoning over FHIR data, which often fails due to complex graph traversal. They developed an RL-enhanced CodeAct agent trained on FHIR-AgentBench, significantly improving answer correctness from 50% to 77%.

Details

FHIR is the dominant standard for the interoperable exchange of healthcare data, representing patient records as a directed graph of resources. Answering clinically meaningful questions requires multi-hop traversal across multiple resource types. However, prior LLM agents often select wrong resources or violate traversal constraints. The authors framed this problem as a sequential decision-making problem. They implemented a multi-turn CodeAct agent and utilized reinforcement learning post-training, guided by an LLM Judge, to update the agent's policy. This end-to-end pipeline significantly improved answer correctness from 50% to 77% on FHIR-AgentBench, demonstrating a reliable method for multi-turn reasoning over structured clinical graphs.

📰
Read Original Article
arxiv.org
→

Original content copyright by respective publishers