Synthetic Hospital - Evidence Retrieval: leaderboard
Synthetic Hospital (Park, Chen and Dettmers, 2026) is an open synthetic longitudinal EHR benchmark built from medical-education material, with labels grounded in ICD-10-CM, SNOMED CT and LOINC. In this task the model ranks chart sections as evidence for a diagnosis, measured by precision at 5 in percent against graded relevance, scored against graph-derived references on the 200-patient public split with one prompting strategy locked per task; higher is better.
Metric: Precision at 5 (%; ranking chart sections for a diagnosis against graded graph relevance, zero-shot; public split of 200 synthetic longitudinal patients (1,268 in all, built from USMLE-style questions with ontology-grounded labels); single-turn, one prompting strategy locked per task for every model; graph-derived deterministic scoring). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.3 | 83.3 | #194 |
| 2 | Claude Opus 4.6 | 81.6 | #100 |
| 3 | Kimi K2.5 (Thinking) | 81.1 | #252 (Kimi K2.5) |
| 4 | Qwen 3.5 397B A17B | 80.9 | #242 |
| 5 | Gemma 3 27B | 80.4 | #1638 |
| 6 | GLM-5 | 79.8 | #256 |
| 7 | DeepSeek V3.2 | 75.1 | #403 |
| 8 | Llama 4 Scout | 74.1 | #1787 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=synthetic-hospital-evidence-retrieval · How It Works · Data refreshed daily, snapshot 2026-11-12.