Med-RewardBench: leaderboard
Medical multimodal judging from Med-RewardBench (Ding et al., 2025): given a medical image, a clinical question and two model responses, the model picks the better response on six dimensions (accuracy, relevance, comprehensiveness, creativity, responsiveness and overall) across 1,026 cases from 13 organ systems. Percent agreement with three physicians' preferences; higher is better, and chance is about 50%.
Metric: Expert-Agreement Accuracy (%; pairwise, six dimensions). Source: arxiv.org. Saturation forecast: Estimated already saturated. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 | 68.86 | #391 |
| 2 | Qwen 2.5 VL 7B | 64.99 | #1796 |
| 3 | Gemini 1.5 Pro | 63.51 | #1053 |
| 4 | InternVL3-8B | 62.3 | #1695 |
| 5 | Pixtral-12B | 61.63 | #2493 |
| 6 | Claude 3.5 Sonnet | 60.26 | #747 |
| 7 | InternVL2-8B | 56.42 | #3025 |
| 8 | MiniCPM-V-2.6 | 55.23 | #3095 |
| 9 | Qwen 2 VL 7B Instruct | 52.16 | #2569 |
| 10 | GPT-4o | 52.06 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=med-rewardbench · How It Works · Data refreshed daily, snapshot 2026-11-12.