UCLA

World Foundation Models (WFMs) are state-of-the-art generative models that aim to simulate dynamics of the physical world across various modalities. While WFMs appear promising for accelerating AI development, there is currently an absence of benchmarks that comprehensively evaluate their physical accuracy. Existing benchmarks are often based on binary object contact prediction, failing to encapsulate the depth and complexity of real-world physics. To address this, we introduce a video-based benchmark that assesses WFMs across a framework of 4 categories: motion physics, object permanence, support relations, and perspective/scale relations. Our dataset comprises 425 physically-accurate scene configurations, plus a language-based subset evaluating the physical reasoning of vision-language models (VLMs) on our generated videos. Our experiments show that leading WFMs and VLMs both lack an accurate understanding of dynamic scenarios. Our benchmark establishes an in-depth approach to evaluate physical reasoning tasks, laying a scalable and generalizable foundation for simulation-driven learning.