CureMed-Bench - Language Consistency: leaderboard
Target-language fidelity on CureMed-Bench (Onyame et al., 2026): the share of final answers written in the language of the question, pooled over 1,431 held-out open-ended medical questions in 13 languages, zero-shot. Scored separately from correctness, in percent; higher is better.
Metric: Language consistency (%; share of final answers written in the language of the question, pooled over 1,431 held-out open-ended medical questions in 13 languages; zero-shot). Source: arxiv.org. Saturation forecast: Estimated already saturated. 49 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3 Haiku | 93.43 | #2629 |
| 2 | Llama 3.3 70B Instruct | 79.66 | #1353 |
| 3 | Llama 3.1 70B | 75.68 | #1493 |
| 4 | GPT-5 Mini | 75.33 | #349 |
| 5 | Qwen 2.5 72B Instruct | 70.73 | #1093 |
| 6 | GPT-5 Nano | 69.11 | #888 |
| 7 | Qwen 3 32B | 55.83 | #960 |
| 8 | Gemini 2.5 Flash | 48.01 | #452 |
| 9 | Llama 3 8B UltraMedical | 47.03 | #6117 |
| 10 | Ministral 8B | 46.93 | #4511 |
| 11 | Qwen 3 14B | 43.71 | #1235 |
| 12 | Qwen 2.5 32B Instruct | 41.51 | #1216 |
| 13 | Llama 3.1 8B Instruct | 36.56 | #4277 |
| 14 | Qwen 2.5 14B Instruct | 35.57 | #1745 |
| 15 | Qwen 3 8B | 33.96 | #1838 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=curemed-bench-language-consistency · How It Works · Data refreshed daily, snapshot 2026-11-12.