MM-JudgeBias - Image Misalignment: 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-Deviation from 0 to 1, the normalized score drop when the image is swapped for an unrelated one; higher is better.

Metric: Bias-Deviation (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 drop of the judge's score, normalized by the largest possible drop, when the image is replaced with an unrelated one; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 26 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.50.973#160
2Gemini 3 ProHigh0.955#145
3Claude Sonnet 4.5Thinking0.905#261
4Gemini 2.5 Pro0.904#276
5Claude Haiku 4.50.818#583
6Gemini 2.5 Flash0.76#463
7Qwen 3 VL 30B A3B Instruct0.648#895
8Gemini 2.5 Flash LiteThinking0.634#964
9Qwen 3 VL 8B Instruct0.622#1029
10Qwen 3 VL 8B (Thinking)0.529#910
11Qwen 3 VL 30B A3B (Thinking)0.476#719
12InternVL3.5-14B0.464#1504
13LLaVA-Critic-7B0.452
14InternVL3.5-30B-A3B0.377#1267
15O3High0.337#237

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