Mind's Eye - Visual Relation Abstraction: leaderboard

Mind's Eye (2026) is a multiple-choice benchmark of visuospatial reasoning tasks adapted from human intelligence tests, with procedurally generated items answered with chain-of-thought prompting. This board is accuracy in percent on a Bongard-style task of finding the figure that lacks the attribute the others share; higher is better.

Metric: Accuracy (%) on the visual relation abstraction (Bongard-style, find the figure lacking the shared attribute) task of Mind's Eye (6-option multiple choice, 100 procedurally generated items, chain-of-thought prompt, answers extracted by an LLM evaluator; keys uniformly distributed over the options); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro29#276
2GPT-4o28.4#751
3Qwen 2.5 VL 32B Instruct25.1#1062
4Idefics-8B (MindsEye checkpoint unspecified)24
5GT-Thinker-7B23.1
6Vision-G1-7B22.3
7Llama 3.2 11B Vision22#4264
8MindsEye InternVL3-8B (checkpoint unspecified)22
9Phi-3.5-Vision Instruct (MindsEye checkpoint unspecified)22
10V-Thinker-8B (MindsEye checkpoint unspecified)21.5
11O321.1#237
12Qwen 2.5 VL 3B Instruct20#2713
13Qwen 2.5 VL 7B Instruct19.1#1902
14InstructBLIP-4B (MindsEye checkpoint unspecified)16.3
15llava-v1.6-mistral-7B16#3999

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

Interactive version: theaggregate.ai/benchmark?slug=mind-s-eye-visual-relation-abstraction · How It Works · Data refreshed daily, snapshot 2026-11-21.