PhyGround: Benchmarking Physical Reasoning in Generative World Models
Researchers have introduced PhyGround, a new benchmark designed to evaluate physical reasoning capabilities in generative world models used for video generation. Addressing limitations in existing benchmarks, such as coarse evaluation frameworks and annotator bias, PhyGround features 250 curated prompts linked to expected physical outcomes across 13 laws of solid-body mechanics, fluid dynamics, and optics. The study involved a large-scale, quality-controlled human experiment with 459 annotators, yielding over 37,400 fine-grained labels with high reliability. Additionally, the team released PhyJudge-9B, an open-source physics-specialized Vision-Language Model judge that demonstrates significantly lower aggregate relative bias compared to Gemini-3.1-Pro. This initiative aims to provide a rigorous, reproducible standard for assessing whether AI-generated videos accurately adhere to real-world physical rules, enhancing the development of more realistic and physically consistent generative models.
Wire timeline
PhyGround: Benchmarking Physical Reasoning in Generative World Models
Researchers have introduced PhyGround, a new benchmark designed to evaluate physical reasoning capabilities in generative world models used for video generation. Addressing limitations in existing benchmarks, such as coarse evaluation frameworks and annotator bias, PhyGround features 250 curated prompts linked to expected physical outcomes across 13 laws of solid-body mechanics, fluid dynamics, and optics. The study involved a large-scale, quality-controlled human experiment with 459 annotators, yielding over 37,400 fine-grained labels with high reliability. Additionally, the team released PhyJudge-9B, an open-source physics-specialized Vision-Language Model judge that demonstrates significantly lower aggregate relative bias compared to Gemini-3.1-Pro. This initiative aims to provide a rigorous, reproducible standard for assessing whether AI-generated videos accurately adhere to real-world physical rules, enhancing the development of more realistic and physically consistent generative models.
cs.AI updates on arXiv.org