WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors

arXiv 2026 65.9 benchmark

TLDR

Introduces WorldReasonBench, a benchmark to test video generators as world-state predictors, revealing a gap between visual plausibility and reasoning.

Reasoning

Strengths: Novel benchmark with structured QA and human-aligned evaluation, covering multiple reasoning dimensions. Weaknesses: Limited to 436 test cases; no discussion of scalability or generalization to diverse domains.

Read-first score

Read-first score 65.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 47.

Recency 6%

100

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

Methodology quality 18%

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

Citation impact 18%

75.1

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.75056323

Reproducibility 18%

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

Topical relevance 29%

67.1

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 429.

Keyword Scores

video world model

world dynamics prediction

world simulator

world model

generative world model

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

WorldReasonBench: a benchmark that reframes video generation evaluation as world-state prediction, testing physical, social, logical, and informational consistency.
Two-part human-aligned evaluation methodology: Process-aware Reasoning Verification (structured QA and reasoning-phase diagnostics) and Multi-dimensional Quality Assessment (reasoning quality, temporal consistency, visual aesthetics).
WorldRewardBench: a preference benchmark with approximately 6K expert-annotated pairs over 1.4K videos for pair-wise and point-wise reward-model evaluation.
Identification of a persistent gap between visual plausibility and world reasoning in modern video generators.

Methodology

WorldReasonBench contains 436 curated test cases with structured ground-truth QA annotations spanning four reasoning dimensions (physical, social, logical, informational) and 22 subcategories. Evaluation uses a two-part methodology: Process-aware Reasoning Verification detects temporal and causal failures via structured QA and reasoning-phase diagnostics, while Multi-dimensional Quality Assessment scores reasoning quality, temporal consistency, and visual aesthetics for ranking and reward modeling. Additionally, WorldRewardBench provides approximately 6K expert-annotated preference pairs over 1.4K videos to support reward-model evaluation.

Key Results

Across modern video generators, results expose a persistent gap between visual plausibility and world reasoning: videos can look convincing while failing dynamics, causality, or information preservation.

Limitations

Benchmark size is limited to 436 test cases, which may not capture the full diversity of real-world scenarios.
Evaluation relies on structured ground-truth QA annotations, which may introduce subjectivity and are costly to scale.
The benchmark covers only four reasoning dimensions (physical, social, logical, informational) and 22 subcategories, potentially missing other aspects of world simulation.

Resources

Paper Code