Switch language한국어
Back to the list

YoCausal: How Far is Video Generation from World Model? A Causality Perspective

TL;DR AI

Key summary

2 min read
  1. YoCausal is a new two-level benchmark for testing causal understanding in video diffusion models.

  2. It uses reversed real-world videos as counterfactuals and measures arrow-of-time perception with the Reverse Surprise Index.

  3. A second metric, the Causality Cognition Index, separates causal from non-causal cases.

  4. Across 13 leading models, the study finds temporal awareness does not imply true causal reasoning, and human performance remains far ahead.

Read the original