XNews

PerceptionBench: 3,000 Questions Isolating 10 Atomic Visual Capabilities from Model Failures

PerceptionBench: 3,000 Questions Isolating 10 Atomic Visual Capabilities from Model Failures

Kimi Moonshot released PerceptionBench, a benchmark that isolates visual perception into 10 atomic capabilities derived from how frontier models fail across 42 existing benchmarks, rather than being predefined. It contains 3,000 verified questions, each designed to test a single perceptual capability and answerable purely by looking, without requiring reasoning or external knowledge. The benchmark aims to provide a granular evaluation of visual perception in AI models.

Source posts