MM-JudgeBias: leaderboard
Metric: Mean judge-reliability score (0 to 1) over nine bias types MM-JudgeBias (about 200 image-query-response triplets per bias type, 1,804 in all, each judged on a 1-10 scale before and after a controlled perturbation; mean of three runs): Bias-Deviation (normalized score drop when the image or query is removed or mismatched) for five types and Bias-Conformity (score invariance under meaning-preserving perturbations) for four; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 30 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Pro | 0.87 |
| 2 | Gemini 3 Pro (High) | 0.87 |
| 3 | Claude Opus 4.5 (Thinking) | 0.86 |
| 4 | Claude Sonnet 4.5 (Thinking) | 0.82 |
| 5 | Claude Opus 4.5 | 0.82 |
| 6 | Gemini 2.5 Flash | 0.76 |
| 7 | Claude Haiku 4.5 (Thinking) | 0.76 |
| 8 | Claude Sonnet 4.5 | 0.75 |
| 9 | Gemini 2.5 Flash Lite (Thinking) | 0.75 |
| 10 | Qwen 3 VL 8B Instruct | 0.74 |
| 11 | Qwen 3 VL 30B A3B (Thinking) | 0.71 |
| 12 | Claude Haiku 4.5 | 0.71 |
| 13 | Qwen 3 VL 30B A3B Instruct | 0.68 |
| 14 | O3 (High) | 0.68 |
| 15 | Gemini 2.5 Flash Lite | 0.67 |
Interactive version: theaggregate.ai/benchmark?slug=mm-judgebias · How It Works · Data refreshed daily, snapshot 2026-10-07.