AI Vision Test Reveals Model Hallucinations

Moonshot AI has launched PerceptionBench, a new AI benchmark designed to rigorously test the visual perception skills of multimodal large language models (MLLMs). This AI benchmark aims to dissect how well these models truly 'see' by focusing on atomic visual capabilities, rather than overall reasoning ability. You can learn more about the work on kimi.com.

Developed by the Kimi Team, PerceptionBench stems from analyzing where current frontier models falter. It isolates 10 distinct perceptual capabilities, including visual relation, counting, attribute recognition, depth perception, localization, comparison, fine-grained recognition, context integration, OCR, and hallucination detection.

The benchmark comprises over 3,000 verified questions, each crafted to be answerable by simple observation, requiring no complex inference or external knowledge. This approach ensures the evaluation focuses purely on perceptual accuracy.

Atomic Capabilities Under Scrutiny

PerceptionBench categorizes failures into distinct perceptual tasks, such as identifying the number of red dots in an image or determining the length of a line segment. For instance, it asks questions like "What is the length of side AC?" with an answer derived directly from the visual data.

Other examples include counting objects like plates on a table or identifying the number of faces of cubes in contact with the ground. The benchmark also tests OCR capabilities with questions like "What is the blue number?".

Crucially, the benchmark highlights a significant issue: many correct answers fail to be reproduced upon repeated questioning. This suggests that current MLLMs may be guessing answers rather than consistently perceiving visual information accurately.

Model Performance Falls Short

The results from evaluating several models are stark. No model tested managed to exceed 60% accuracy on PerceptionBench. Furthermore, models with similar overall scores often displayed vastly different strengths and weaknesses across the 10 perceptual categories.

The Kimi Team's initiative, which also includes advancements like Kimi K2.6, aims to provide a sharp diagnostic tool. This is in contrast to benchmarks that offer a single score, potentially masking underlying perceptual deficits.

The benchmark's design, driven by observed model failures across over 40 existing benchmarks, ensures it targets genuine weaknesses. This failure-driven taxonomy, combined with rigorous verification, makes PerceptionBench a critical tool for advancing faithful and consistent visual perception in multimodal AI.