New benchmark confirms AI models still perform poorly at visual perception

Moonshot AI has introduced PerceptionBench, a benchmark designed to evaluate the visual perception capabilities of multimodal AI models, distinct from their logical reasoning abilities. The findings indicate that no leading model, including the recently highlighted GPT-5.6 Sol, achieves an accuracy exceeding 60%. This suggests a significant gap in the visual processing capabilities of current AI systems.

The research highlights that many errors attributed to reasoning in these models may actually originate during the initial image-reading phase. This insight underscores the importance of improving foundational visual perception in AI, as it appears to be a critical bottleneck affecting overall performance. The results from PerceptionBench raise questions about the reliability of AI models in tasks requiring visual understanding, emphasizing the need for further advancements in this area. For more details, refer to the original article on The Decoder.