M-JudgeBench - Incidental Mistake Detection: leaderboard
Metric: pairwise accuracy (%) on the 252 pairs where the rejected chain has an injected incidental slip (spelling, transcription, units) but the same correct final answer, M-JudgeBench (3,712 multimodal judge pairs from MMMU, MMMU-Pro, MMStar, MMReason, M3CoT, MathVision and MathVerse; the model sees the image, question and two responses and outputs its choice directly, open-source judge models with their own prompts); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview 06-05) | 92.46 | #101 |
| 2 | GPT-4.1 (2025-04-14) | 83.73 | #244 |
| 3 | GPT-5 | 82.14 | #91 |
| 4 | GPT-4.1 Mini | 81.75 | #346 |
| 5 | Qwen 3 VL 8B Instruct | 76.59 | #401 |
| 6 | Qwen 3 VL 4B Instruct | 73.81 | #506 |
| 7 | GPT-4o (2024-08-06) | 70.24 | #326 |
| 8 | GLM-4.5V (Non-reasoning) | 68.25 | #339 (GLM-4.5V) |
| 9 | Qwen 2.5 VL 7B Instruct | 41.27 | #643 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=m-judgebench-incidental-mistake-detection · How It Works · Data refreshed daily, snapshot 2026-10-11.