MM-JudgeBias - Unnecessary Image: 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. Bias-Conformity from 0 to 1, the stability of the score when an unrelated image is added to a text-only task; higher is better.

Metric: Bias-Conformity (0 to 1) on 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): mean score invariance (1 minus the absolute score change over its largest possible value) when an unrelated image is added to a text-only task; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1LLaVA-Critic-72B0.974
2GPT-4.1 Mini0.972#789
3GPT-5.1High0.961#240
4Gemini 2.0 Flash Lite0.958#1050
5GPT-5 MiniHigh0.954#364
6Qwen 3 VL 8B Instruct0.952#1029
7O4 MiniHigh0.95#368
8Gemini 3 ProHigh0.947#145
9Gemini 2.5 Flash0.944#463
10Claude Haiku 4.5Thinking0.943#583
11Gemini 2.5 Flash LiteThinking0.94#964
12O3High0.937#237
13LLaVA-Critic-7B0.929
14Qwen 3 VL 8B (Thinking)0.927#910
15Claude Opus 4.5Thinking0.925#160

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-unnecessary-image · How It Works · Data refreshed daily, snapshot 2026-11-21.