MM-JudgeBias: leaderboard

MM-JudgeBias (2026) tests multimodal LLM judges for compositional bias: the judge scores image, query and response triplets from 1 to 10 before and after a controlled perturbation, and should lower its score when evidence is removed or mismatched and keep it when the change preserves meaning. Mean reliability from 0 to 1 over nine bias types; higher is better.

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 February 2027. 26 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 2.5 Pro0.869#276
2Gemini 3 ProHigh0.869#145
3Claude Opus 4.5Thinking0.858#160
4Claude Sonnet 4.5Thinking0.816#261
5Gemini 2.5 Flash0.761#463
6Claude Haiku 4.5Thinking0.757#583
7Gemini 2.5 Flash LiteThinking0.747#964
8Qwen 3 VL 8B Instruct0.738#1029
9Qwen 3 VL 30B A3B (Thinking)0.713#719
10Qwen 3 VL 30B A3B Instruct0.678#895
11O3High0.675#237
12Qwen 3 VL 8B (Thinking)0.655#910
13O4 MiniHigh0.654#368
14Gemini 2.0 Flash Lite0.647#1050
15LLaVA-Critic-7B0.647

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mm-judgebias · How It Works · Data refreshed daily, snapshot 2026-11-21.