LLMEval-Med: leaderboard
Overall usability on LLMEval-Med's 677-question test set of mostly Chinese, open-ended clinical questions built from electronic health records and expert-designed scenarios: the mean of five category usability rates (medical knowledge, language understanding, reasoning, safety and ethics, text generation). A response counts as usable when a GPT-4o judge using expert checklists averages 4 or more of 5, while text generation is physician-rated; percent, higher is better. Zhang et al. (Fudan University), Findings of EMNLP 2025.
Metric: Overall usability: mean of the five category usability rates (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 | 64.23 | #467 |
| 2 | O1 Preview | 61.23 | #483 |
| 3 | GPT-4o | 58.73 | #741 |
| 4 | O1 Mini | 57.86 | #761 |
| 5 | DeepSeek V3 | 48.03 | #705 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=llmeval-med · How It Works · Data refreshed daily, snapshot 2026-11-12.