VLM-RobustBench (MMMU-Pro): leaderboard
Metric: Mean relative corruption error (percent of visual gain lost): for each of the 133 corrupted settings: 42 severity-based corruptions in nine categories (noise, blur, weather, digital, geometric and others) at low, mid and high severity, plus 7 binary transforms such as vertical flip and colour inversion; direct-answer prompting, the accuracy drop from clean images divided by the visual gain (clean accuracy minus the no-image accuracy), averaged over all 133 settings, on a stratified 20 percent sample (532 questions) of the 10-option MMMU-Pro standard set; 100 means the corruption removes all visual benefit, values can be negative or exceed 100; lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 4B Instruct | -6.8 | #506 |
| 2 | Molmo2-8B | 1 | #522 |
| 3 | Qwen 3 VL 8B Instruct | 4.7 | #401 |
| 4 | Qwen 3 VL 30B A3B Instruct | 12.1 | #365 |
| 5 | Gemma 3 12B (IT) | 24.2 | #655 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=vlm-robustbench-mmmu-pro · How It Works · Data refreshed daily, snapshot 2026-10-11.