Thinking in Video: Can Video Generators Really Reason About the Real World?

arXiv 2026 47.4 benchmark

TLDR

Proposes Causal-Generative Dual-Judge (CGDJ) to evaluate if video generators truly reason about real-world dynamics, revealing a perception-prediction gap.

Reasoning

The paper introduces a novel evaluation framework (CGDJ) that separates explicit causal perception from implicit generative consistency, identifying a key limitation in current video generators. Its strength lies in the dual-judge approach and empirical findings, but it is limited to specific models and does not fully address generalization across diverse real-world scenarios.

Read-first score

Read-first score 47.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 45.

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=analysis,evaluation,experiment,metric

Topical relevance 29%

64.3

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=artifact

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 590.

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

Redefining the video generation reasoning paradigm as 'Thinking in Video', where video serves as a medium for constructing, extending, and verifying causal thought.
Introducing the Causal-Generative Dual-Judge (CGDJ) evaluation framework that audits World Model Consistency via Explicit Causal Perception (spatio-temporal visual question answering) and Implicit Generative Perception-Prediction Gap (future video consistency).
Identifying the Perception-Prediction Gap and audio-visual misalignment, where models verbalize correct causal logic more reliably than they render it visually.

Methodology

The paper proposes CGDJ, a dual-judge framework that evaluates video generators' reasoning by testing their ability to answer causal questions about a video scenario (Explicit Causal Perception) and to generate a consistent future video that matches the causal consequence (Implicit Generative Perception-Prediction Gap). Experiments are conducted on representative open- and closed-source video generators.

Key Results

Open-source video generators exhibit plausible dynamics but near-zero explicit causal perception, while advanced closed-source models show stronger but still limited alignment between reasoning and generation; audio-visual misalignment reveals that models verbalize correct causal logic better than they render it visually.

Limitations

Open-source video generators lack explicit causal perception despite generating plausible dynamics.
Even advanced closed-source models demonstrate only limited alignment between reasoning and generation.
Audio-visual misalignment: models can verbalize correct causal logic more reliably than they can render it in video, challenging the world simulator narrative.

Resources

Paper