{"paper":{"title":"Reinforcement Learning for Tool-Calling Agents in Fast Healthcare Interoperability Resources (FHIR)","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Reinforcement learning post-training raises a Qwen3-8B model to 77 percent correctness on FHIR clinical queries, surpassing larger closed models.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Jan P. Bremer, Marius S. Knorr, Nils Schweingruber, Robert M\\\"uller","submitted_at":"2026-05-13T21:27:21Z","abstract_excerpt":"Fast Healthcare Interoperability Resources (FHIR) is the dominant standard for interoperable exchange of healthcare data. In FHIR, electronic health records form a directed graph of resources. Answering clinically meaningful questions over FHIR requires agents to perform multi-step reasoning, filtering, and aggregation across multiple resource types. Prior work shows that even tool-augmented LLM agents (retrieval, code execution, multi-turn planning) often select the wrong resources or violate traversal constraints. We study this problem in the context of FHIR-AgentBench, a benchmark for reali"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Empirically, our approach improves answer correctness from 50% (o4-mini) to 77% on FHIR-AgentBench using a smaller and cheaper Qwen3-8B model.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That an LLM Judge can supply reliable execution-grounded rewards without bias or systematic errors in judging multi-step FHIR traversals.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"RL post-training lifts answer correctness on FHIR-AgentBench from 50% (o4-mini) to 77% with a cheaper Qwen3-8B CodeAct agent.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning post-training raises a Qwen3-8B model to 77 percent correctness on FHIR clinical queries, surpassing larger closed models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2095daef2331626e1d156c9ec4667e7e43672ac467fb9794f2adbff6b352a9c9"},"source":{"id":"2605.14126","kind":"arxiv","version":1},"verdict":{"id":"2f1d9550-d38b-4e8b-8e77-f703c37219b7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T04:53:35.494733Z","strongest_claim":"Empirically, our approach improves answer correctness from 50% (o4-mini) to 77% on FHIR-AgentBench using a smaller and cheaper Qwen3-8B model.","one_line_summary":"RL post-training lifts answer correctness on FHIR-AgentBench from 50% (o4-mini) to 77% with a cheaper Qwen3-8B CodeAct agent.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That an LLM Judge can supply reliable execution-grounded rewards without bias or systematic errors in judging multi-step FHIR traversals.","pith_extraction_headline":"Reinforcement learning post-training raises a Qwen3-8B model to 77 percent correctness on FHIR clinical queries, surpassing larger closed models."},"references":{"count":66,"sample":[{"doi":"10.1093/jamia/ocx080","year":2017,"title":"Journal of the American Medical Informatics Association , author =","work_id":"c24070b8-a397-4077-97ed-5d4542edbc17","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.5009/gnl230272","year":2024,"title":"Gut and Liver , author =","work_id":"eabf3f84-9d5d-4b21-80a0-54e9e513f7b6","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1001/journalofethics.2017.19.3.stas1-1703","year":2017,"title":"AMA journal of ethics , author =","work_id":"c8fd5c90-8a64-448f-b22f-64ef902ae287","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1002/wics.1549","year":2021,"title":"WIREs Computational Statistics , author =","work_id":"2dfbcab8-e428-4f15-aedd-905ef9aa0e76","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1016/j.isci.2024.109713","year":2024,"title":"iScience , author =","work_id":"acf0f2f5-0d21-4841-94a0-9e376fa33c40","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":66,"snapshot_sha256":"febf422a970a46c5f57efe1720c3f691045ab70d9445ee05ce14580ddab0f610","internal_anchors":8},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}