{"paper":{"title":"RxEval: A Prescription-Level Benchmark for Evaluating LLM Medication Recommendation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"RxEval tests LLMs on specific medication-dose-route choices from detailed patient trajectories, revealing top models reach only 46 percent exact match.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Changmiao Wang, James T. Kwok, Shuhao Chen, Weisen Jiang, Xiaoqing Wu, Xuanren Shi, Yu Zhang","submitted_at":"2026-05-14T08:24:03Z","abstract_excerpt":"Inpatient medication recommendation requires clinicians to repeatedly select specific medications, doses, and routes as a patient's condition evolves. Existing benchmarks formulate this task as admission-level prediction over coarse drug codes with multi-hot diagnostic and procedure code inputs, failing to capture the per-timepoint, information-rich nature of real prescribing. We propose RxEval, a prescription-level benchmark that evaluates LLM prescribing capability by multiple-choice questions: each question presents a detailed patient profile and time-ordered clinical trajectory, requiring "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Evaluation of 16 LLMs shows that RxEval is both challenging and discriminative: F1 ranges from 45.18 to 77.10 across models, and the best Exact Match is only 46.10%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The distractors generated by reasoning-chain perturbation are clinically valid and representative of the kinds of errors or alternatives that would actually arise in real prescribing decisions.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"RxEval benchmark shows frontier LLMs reach at most 46.10% exact match on prescription-level medication, dose, and route selection from real patient trajectories.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"RxEval tests LLMs on specific medication-dose-route choices from detailed patient trajectories, revealing top models reach only 46 percent exact match.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"17d574d17628c75d8ef06a445f1d93a1bad856d7974a84f1acb947471e7a8d4d"},"source":{"id":"2605.14543","kind":"arxiv","version":1},"verdict":{"id":"29889642-0f2e-434a-8e13-0d5b65f6a1f4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:04:49.937450Z","strongest_claim":"Evaluation of 16 LLMs shows that RxEval is both challenging and discriminative: F1 ranges from 45.18 to 77.10 across models, and the best Exact Match is only 46.10%.","one_line_summary":"RxEval benchmark shows frontier LLMs reach at most 46.10% exact match on prescription-level medication, dose, and route selection from real patient trajectories.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The distractors generated by reasoning-chain perturbation are clinically valid and representative of the kinds of errors or alternatives that would actually arise in real prescribing decisions.","pith_extraction_headline":"RxEval tests LLMs on specific medication-dose-route choices from detailed patient trajectories, revealing top models reach only 46 percent exact match."},"references":{"count":54,"sample":[{"doi":"","year":null,"title":"V ALS MedQA leaderboard. Website. URLhttps://www.vals.ai/benchmarks/medqa","work_id":"61cbc542-3336-4a24-bd05-75d81dec0e07","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"K. Agyeman-Manu, T. A. Ghebreyesus, M. Maait, A. Rafila, L. Tom, N. T. Lima, and D. Wangmo. Prioritising the health and care workforce shortage: Protect, invest, together.The Lancet Global Health, 11 ","work_id":"a6d621d1-aff2-4deb-8f4d-a299611274e9","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Z. Ali, Y . Huang, I. Ullah, J. Feng, C. Deng, N. Thierry, A. Khan, A. U. Jan, X. Shen, W. Rui, et al. Deep learning for medication recommendation: A systematic survey.Data Intelligence, 5(2):303–354,","work_id":"33b4fbfd-d5ca-4fa4-aaa8-b9af82c854f1","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Qwen2.5 Technical Report","work_id":"d8432992-4980-4a81-85c7-9fa2c2b87f85","ref_index":4,"cited_arxiv_id":"2412.15115","is_internal_anchor":true},{"doi":"","year":2025,"title":"HealthBench: Evaluating Large Language Models Towards Improved Human Health","work_id":"5d613c2c-158f-488c-9981-fc9b82a5e093","ref_index":5,"cited_arxiv_id":"2505.08775","is_internal_anchor":true}],"resolved_work":54,"snapshot_sha256":"71532895065a8cf02743e66cce5d543ddf63822f924cab13e0f895b3f8f9b555","internal_anchors":12},"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"}