Atlas

Benchmarks

← All benchmarks

PerceptionBench - Visual Relation

Multimodal · 2026-07-16

The PerceptionBench visual-relation category contains 330 questions that test whether a model can perceive visible spatial, directional, and structural relationships between image elements.

Top models (higher is better)

ModelScore
GPT-5.6 Sol69.7
Kimi K368.2
GPT-5.561.9
Qwen3.7-Plus59.1
Gemini 3.1 Pro Preview58.8
Claude Fable 558.5
Doubao Seed 2.1 Pro57.6
Qwen3.5 397B A17B55.2
Gemini 3.5 Flash53.6
Opus 4.851.4
Kimi K2.650.9
Grok 4.547.0
Gemma 4 31B IT42.7
GLM 5V Turbo41.2
MiniMax M340.0
GLM-4.6V35.2
Loading Atlas data…