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

arXiv 2026 41.5 method

TLDR

Proposes Causal-Generative Dual-Judge to evaluate whether video generators reason about real-world dynamics, revealing a perception-prediction gap between causal understanding and generated outcomes.

Reasoning

Strengths: introduces a dual evaluation framework separating explicit causal perception from implicit generative prediction, and applies it to representative video generators. Weaknesses: the abstract lacks detailed methodology and quantitative results, and the approach may not fully resolve concerns about memorization versus causal understanding.

Read-first score

Read-first score 41.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 5%

100

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

Methodology quality 15%

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

Topical relevance 25%

68.6

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 15%

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

Graph prestige 15%

10.9

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0031

Citation impact 15%

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

Citation velocity 10%

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: 496.

Keyword Scores

world model

generative world model

video world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Defines and formalizes the 'Thinking in Video' paradigm, treating video as a medium for constructing, extending, and verifying causal thought rather than merely an output artifact.
Introduces the Causal-Generative Dual-Judge (CGDJ) to audit World Model Consistency through two complementary evaluations: Explicit Causal Perception and Implicit Generative Perception-Prediction Gap.
Empirically identifies a Perception-Prediction Gap and audio-visual misalignment in representative open- and closed-source video generators.

Methodology

The paper proposes CGDJ, which audits World Model Consistency from two perspectives: Explicit Causal Perception uses spatio-temporal flattened visual question answering to test whether a generator interprets a video scenario as a reasoning problem, while Implicit Generative Perception-Prediction Gap evaluates whether the generator renders the causal consequence as a consistent future video. The method is applied to representative open- and closed-source video generators.

Key Results

Open-source models produce plausible dynamics despite near-zero explicit causal perception, whereas advanced closed-source systems show stronger but still limited alignment between reasoning and generation. Analysis also exposes audio-visual misalignment, with models verbalizing correct causal logic more reliably than they render it.

Limitations

Convincing video rollouts may reflect memorized appearances rather than causal understanding.
Existing metrics separate perceptual fidelity from semantic logic, motivating but also highlighting the difficulty of unified evaluation.
Open-source generators exhibit near-zero explicit causal perception.
Advanced closed-source systems still show limited alignment between reasoning and generation.
Models exhibit audio-visual misalignment, challenging the 'world simulator' narrative.

Resources

Paper