This research introduces HVSBench, a benchmark for assessing how closely Multimodal Large Language Models perceive visual information like the human visual system.

• Evaluates MLLMs on fundamental visual perception abilities including attention allocation, visual illusions, and cognitive biases
• Reveals significant gaps between human and MLLM visual perception capabilities
• Provides insights for developing more human-aligned visual AI systems

For security applications, understanding these perception gaps is crucial when deploying vision-based AI in monitoring systems, as misalignments could lead to critical oversights or false detections that humans wouldn't make.

Do Multimodal Large Language Models See Like Humans?