Vision-Language Models (VLMs) have demonstrated strong multimodal reasoning abilities, but it remains unclear whether they genuinely grasp the physical consistency that governs real-world dynamics. According to a new research paper, existing evaluation paradigms tend to be fragmented, examining isolated cognitive stages while neglecting the interplay between perception, reasoning, and physical judgment.
To address this, the researchers introduce PhysVista, a benchmark designed to assess physical intelligence in VLMs through a closed cognitive loop inspired by the human process of seeing, reasoning, and assessing. The benchmark jointly evaluates three components: physical state perception, physical dynamics reasoning, and physical plausibility assessment.
PhysVista further distinguishes between event-level reasoning and scale-level reasoning, enabling fine-grained analysis of how models understand physical phenomena. It also incorporates both real-world and AI-generated videos, allowing evaluation across diverse domains and emerging generative scenarios.
Extensive experiments across a range of VLMs reveal substantial limitations in physical reasoning and plausibility assessment. The findings highlight a persistent gap between visual recognition and genuine physical understanding, and point toward more principled designs for physically grounded multimodal intelligence.