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

arXiv 2026 57.3 benchmark

TLDR

YoCausal benchmarks video diffusion models' causal understanding using reversed real-world videos, revealing a gap between temporal perception and true causality.

Reasoning

The paper introduces a novel benchmark (YoCausal) with two levels to disentangle temporal bias from causal reasoning, using real-world videos and cognitive science principles. Strengths include a zero-cost counterfactual generation method and evaluation of 13 models; weaknesses are that the benchmark's scalability and human-level comparison details are not fully elaborated in the abstract.

Read-first score

Read-first score 57.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 40.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

88.3

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.88296748

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,dataset,evaluation

Topical relevance 29%

57.1

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 352.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

interactive world model

Deep Analysis

Innovations

Two-level benchmark inspired by the Violation of Expectation (VoE) paradigm from cognitive science
Reverse Surprise Index (RSI) for quantifying arrow-of-time perception via denoising loss
Causality Cognition Index (CCI) using a VLM to stratify datasets into causal and non-causal subsets
Zero-cost generation of natural counterfactual samples by temporally reversing real-world videos

Methodology

YoCausal is a two-level benchmark that evaluates video diffusion models on causal understanding. Level 1 uses the Reverse Surprise Index (RSI), which measures arrow-of-time perception by comparing denoising losses on original and temporally reversed videos. Level 2 introduces the Causality Cognition Index (CCI), which leverages a vision-language model (VLM) to partition datasets into causal and non-causal subsets, enabling disentanglement of genuine causal reasoning from temporal bias. The benchmark is arbitrarily extensible as it uses real-world videos reversed at zero cost.

Key Results

Evaluation of 13 state-of-the-art video diffusion models reveals that perceiving the arrow of time does not imply understanding causality, and a significant gap persists relative to human-level causal cognition.

Resources

Paper