DiCoBench - Attribute Difference: leaderboard

Metric: Accuracy (%) on the attribute-difference task (a subtle change of colour, material or shape of one object) of DiCoBench (765 multi-image samples at near-2K resolution): 5-option multiple choice (four candidate image pairs or instructions plus a no-visible-difference option, chance 20%), answer letter matched exactly, VLMEvalKit settings at temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 18 models tracked.

Top models

#ModelScore
1Gemini 3 Pro49.4
2GPT-4.1 Mini42.4
3O4 Mini38.8
4GPT-535.3
5Qwen 2.5 VL 7B Instruct33
6GPT-4o32.9
7GPT-4.128.3
8Qwen 3.5 35B A3B28.3
9Gemini 3 Flash25
10Gemma 3 12B (IT)22.4
11Qwen 2.5 VL 32B Instruct22.4
12Gemma 3 27B (IT)20

Interactive version: theaggregate.ai/benchmark?slug=dicobench-attribute-difference · How It Works · Data refreshed daily, snapshot 2026-09-29.