MR-Bench: leaderboard
Metric: Accuracy (%, times 100), mean of the medication-imputation and procedure-selection tasks; 1,000 MIMIC-IV hospital admissions, eight-option questions, temperature 0.5, an answer counted from the first or the last option letter the output names, whichever scores higher over the dataset; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 24 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 64.1 |
| 2 | Gemini 3 Pro | 61.3 |
| 3 | DeepSeek V3.2 (Thinking) | 58.9 |
| 4 | Qwen 3 Max | 56.4 |
| 5 | GPT-4o | 50.6 |
| 6 | Qwen 3 8B | 43.6 |
| 7 | Qwen 3 4B 2507 (Thinking) | 43 |
| 8 | Qwen 2.5 7B Instruct | 40.4 |
| 9 | Gemma 3 4B (IT) | 37.1 |
| 10 | Llama 3 8B Instruct | 29.9 |
| 11 | Llama 3.1 8B Instruct | 28.5 |
| 12 | MedGemma-4B | 26.9 |
Interactive version: theaggregate.ai/benchmark?slug=mr-bench · How It Works · Data refreshed daily, snapshot 2026-10-07.