MMPCBench - Error Detection Accuracy: leaderboard
MMPCBench (arXiv 2608.29286v1, Table 2): proactive critique of flawed multimodal inputs for 14 MLLMs; a judge panel of GPT-5.2, Gemini-3-Pro and Claude-Sonnet-4 scores the responses, and two of those judges are also evaluated. Share of flawed inputs whose defect the model flags without being prompted to look for one; percent, higher is better.
Metric: Accuracy (%; flaw flagged without prompting). Source: arxiv.org. Saturation forecast: Around 2032. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 235B A22B Instruct | 60 | #280 |
| 2 | Qwen 3 VL 30B A3B Instruct | 55 | #390 |
| 3 | Qwen 3 VL 8B Instruct | 53.3 | #419 |
| 4 | Claude Sonnet 4.5 (Thinking) | 40.2 | #148 (Claude Sonnet 4.5) |
| 5 | Gemini 3 Pro | 37.4 | #82 |
| 6 | Gemini 3 Flash | 32.3 | #103 |
| 7 | GPT-5.1 (High) | 31.7 | #145 (GPT-5.1) |
| 8 | GPT-5.2 | 28.2 | #115 |
| 9 | Gemma 3 12B (IT) | 27 | #672 |
| 10 | Gemma 3 27B (IT) | 25.9 | #528 |
| 11 | GPT-5 Mini | 24.8 | #193 |
| 12 | Gemma 3 4B (IT) | 18.6 | #997 |
| 13 | GLM-4.6V | 16 | #330 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mmpcbench-error-detection-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.