Benchmark RadarYoCausal
YoCausal is a two-level benchmark that evaluates video diffusion models' understanding of temporal causality using the Violation of Expectation paradigm. It temporally reverses real-world videos as counterfactual samples and introduces two metrics: the Reverse Surprise Index (RSI) for arrow-of-time perception and the Causality Cognition Index (CCI) for disentangling genuine causal reasoning from temporal bias.
- Released
- 2026-05-28
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Existing video benchmarks rely on synthetic data and do not separate temporal-direction awareness from true causal understanding. YoCausal provides a protocol that isolates causal cognition from temporal bias, enabling evaluation of whether models genuinely understand cause-and-effect relationships or only statistical patterns.
Motivation
As video diffusion models (VDMs) advance toward world models, a key question arises: do they truly understand causality, or merely overfit to statistical temporal patterns?
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.