MMErroR: leaderboard
MMErroR (Shi et al., ACL 2026) error type classification: given an image, a question and a reasoning chain containing exactly one injected error, a vision-language model names the error type (visual perception, knowledge deployment, question comprehension or reasoning). Sample-weighted accuracy (%) over 1,997 samples in six domains; higher is better.
Metric: Error Type Classification Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 66.65 | #111 |
| 2 | Seed 2.0 Pro | 64.8 | #209 |
| 3 | GPT-5.2 (xHigh) | 64.3 | #172 (GPT-5.2) |
| 4 | Claude Opus 4.5 | 61.04 | #152 |
| 5 | Kimi K2.5 | 60.19 | #252 |
| 6 | Grok 4 | 60.19 | #306 |
| 7 | Qwen 3 VL 32B (Thinking) | 59.29 | #557 |
| 8 | Qwen-VL-Max | 58.19 | #1226 |
| 9 | GLM-4.6V | 50.23 | #698 |
| 10 | Qwen 3 VL 32B Instruct | 43.01 | #582 |
| 11 | Llama 4 Maverick Instruct | 39.46 | #1125 |
| 12 | Llama 4 Scout Instruct | 39.06 | #1532 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mmerror · How It Works · Data refreshed daily, snapshot 2026-11-12.