Apple-π Benchmark Tests Whether Video Models Reason Through Physical Laws, Not Just Mimic ThemApple-π基准发布：检验视频生成模型是否真正依据物理定律推理，而非表面模仿
S 1.7T11 sources1 个来源R7-research
A 14-author team led by Runmao Yao (with Ziwei Liu as senior author) introduced Apple-π (arXiv:2607.16401, submitted 17 Jul 2026), the first benchmark designed to check whether video generation models reach physically plausible results via a faithful, law-grounded reasoning process rather than being judged only on output plausibility.
Apple-π is built from three components, the first being Orchard, a dataset of 400 videos constructed to probe models' grounding in physical laws.
As video generation models are increasingly promoted as emerging "world models," a benchmark that scrutinizes law-grounded reasoning rather than surface-level plausibility bears directly on autonomous-driving simulation, where physics-faithful world models are needed to generate reliable scenarios for AV testing and validation.