YoCausal: How Far is Video Generation from World Model? A Causality Perspective
YoCausal introduces a two-level benchmark to test if video generation models understand causality or just temporal patterns. Level 1, the Reverse Surprise Index (RSI), measures perception of time's arrow. Level 2, the Causality Cognition Index (CCI), separates genuine causal reasoning from simple temporal bias using a Vision-Language Model. Findings show models perceive time but lack true causal understanding.