PerceptionBench Evaluates Atomic Visual Perception in MLLMs
July 26, 2026
PerceptionBench provides a bottom-up evaluation of Multimodal Large Language Models using 3,000 verified questions across ten atomic perceptual capabilities. It isolates visual perception errors from reasoning or domain knowledge failures to diagnose specific model weaknesses.
HOW THIS AFFECTS YOU
●
researcherYou can use this to isolate whether MLLM failures stem from perception or high-level reasoning.