The year/Independent research

Paper 2605.30346

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

Published
May 2026
Research lab
Independent
Citations
1
GitHub
36 stars

01 In brief

Summary

YoCausal is a two-level benchmark for evaluating causal cognition in video diffusion models (VDMs), inspired by the Violation of Expectation paradigm.

It uses temporally reversed real-world videos as counterfactual samples, avoiding synthetic data and the sim-to-real gap.

Level 1 introduces the Reverse Surprise Index (RSI), measuring arrow-of-time perception via denoising loss.

Level 2 introduces the Causality Cognition Index (CCI), which uses a VLM to split videos into causal and non-causal subsets, isolating genuine causal reasoning from temporal bias.

Evaluation of 13 state-of-the-art VDMs shows that perceiving the arrow of time does not imply understanding causality, and a significant gap remains compared to human-level causal cognition.

The benchmark is extensible, using 1,232 real-world videos across four domains, and includes a human upper bound.

Key findings include that causal cognition correlates with intuitive physics but not aesthetic quality, and that scaling parameters and architectural evolution improve causal cognition.

02 From the paper

Abstract

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? Existing benchmarks mostly rely on synthetic data, limiting real-world generalization due to the sim-to-real gap. We present YoCausal, a two-level benchmark inspired by the Violation of Expectation (VoE) paradigm from cognitive science. By temporally reversing real-world videos at zero cost as natural counterfactual samples, YoCausal establishes an arbitrarily extensible evaluation protocol. Level 1 introduces the Reverse Surprise Index (RSI), quantifying arrow-of-time perception via denoising loss. Level 2 introduces the Causality Cognition Index (CCI), which leverages a VLM to stratify datasets into causal and non-causal subsets, disentangling genuine causal reasoning from temporal bias. Evaluation of 13 state-of-the-art VDMs reveals that perceiving the arrow of time does not imply understanding causality, and a significant gap persists relative to human-level causal cognition.