Research map / Multi-agent systems and games / LLM agents

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

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Post-trains Qwen3-8B as a multi-turn CodeAct FHIR QA agent with GRPO and an LLM-judge reward on FHIR-AgentBench, raising answer correctness from about 20% to 77% and beating o4-mini (47%) and other closed baselines.

Problem

Answering clinical questions over FHIR resource graphs requires multi-hop traversal, filtering, and aggregation, and open-weight LLM agents underperform closed models on this realistic EHR tool-use task.

Approach

Frame FHIR question answering as sequential decision-making over a resource graph and apply execution-grounded RL (GRPO with an LLM-judge binary reward) to a multi-turn CodeAct agent with retrieval and Python tools, with dynamic sampling that filters zero-variance groups.

Key results

• Qwen3-8B + RL (GRPO + clip-higher + DAPO filter) reaches 76.7% ± 0.9 answer correctness on FHIR-AgentBench validation, versus 19.7% for vanilla Qwen3-8B and 46.9% for o4-mini.
• Judge agreement with a board-certified physician is 96.2% (precision 96.7%, recall 97.4%).
• Symmetric vs. asymmetric clipping give nearly identical results (64% vs. 62%), while adding DAPO dynamic sampling reaches 77% in fewer steps.

Surprising findings

• RL-trained agent learns to answer within a six-turn budget while vanilla Qwen3-8B uses a 12-turn budget, so accuracy rises while turns decrease.
• The agent first learns to return empty answers (cheapest reward) before improving on data-retrieval question types, with some negative transfer to the Empty category.

Limitations

• Retrieval tool fetches whole resource types and will not scale to large patient populations without query planning.
• Evaluation relies on an LLM judge with 96% agreement with one physician annotator; 77% accuracy remains far below clinical reliability.
• Failures concentrate on multi-hop reference traversal, e.g. MedicationRequest-Medication resolution, applied inconsistently.

Experimental setup

Datasets: FHIR-AgentBench MIMIC-IV FHIR demo (100 patients)

Benchmarks: FHIR-AgentBench

Metrics: answer correctness pass@5 per-turn score per-resource-type accuracy

Baselines: o4-mini Gemini 3 Flash GLM-5 Qwen3-4B Qwen3-8B Qwen3-14B Qwen3-32B

Builds on: GRPO CodeAct DAPO SkyRL FHIR-AgentBench

Closest papers by meaning

• MedChain: Bridging the Gap Between LLM Agents and Clinical Practice with Interactive Sequence (NeurIPS 2025 · similarity 0.87)
• MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning (ICLR 2026 · similarity 0.87)
• MedLA: A Logic-Driven Multi-Agent Framework for Complex Medical Reasoning with Large Language Models (AAAI 2026 · similarity 0.87)
• AGILE: A Novel Reinforcement Learning Framework of LLM Agents (NeurIPS 2024 · similarity 0.86)
• Doctor-R1: Mastering Clinical Inquiry with Experiential Agentic Reinforcement Learning (ICLR 2026 · similarity 0.86)
• MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments (NeurIPS 2026 · similarity 0.86)