Atlas

Benchmarks

← All benchmarks

PerceptionBench - Hallucination

Multimodal · 2026-07-16

The PerceptionBench hallucination category contains 271 questions that test whether a model avoids asserting plausible but nonexistent objects, text, or attributes when the visual evidence is absent.

Top models (higher is better)

ModelScore
Gemini 3.5 Flash50.6
Doubao Seed 2.1 Pro49.8
Claude Fable 545.0
Grok 4.544.7
Kimi K341.7
Kimi K2.641.0
Gemini 3.1 Pro Preview40.6
Opus 4.838.7
GPT-5.534.7
Gemma 4 31B IT32.1
MiniMax M329.9
Qwen3.7-Plus29.5
GLM 5V Turbo28.0
GLM-4.6V26.9
GPT-5.6 Sol26.9
Qwen3.5 397B A17B26.9
Loading Atlas data…