Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning

arXiv 2026 44.8 method

TLDR

Introduces Latent Dynamics Reasoning, a video world model that extrapolates physics dynamics from pixels via kinematic integration, outperforming diffusion baselines on white-box benchmarks.

Reasoning

The paper proposes a novel latent dynamics integration method for video world models, with strong extrapolation results and efficiency gains over diffusion baselines. However, validation is limited to a controlled synthetic physics benchmark, and the abstract does not demonstrate real-world or interactive environments.

Read-first score

Read-first score 44.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

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=baseline,benchmark,evaluation

Topical relevance 29%

61.4

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=github

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: 472.

Keyword Scores

video world model

world model

world dynamics prediction

generative world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Latent Dynamics Reasoning (LDR): models latent transitions as explicit kinematic integration, numerically integrating lower-order dynamics and learning only a third- and higher-order residual to drive the rollout.
Use of a structured latent space instead of dense convolutional features to improve out-of-distribution extrapolation of dynamics.
First video world model that extrapolates learned dynamics beyond its training distribution, demonstrated by severe shift generalization (e.g., red balls to blue square).

Methodology

LDR casts video prediction as an explicit kinematic integration in a structured latent space, where lower-order dynamics are numerically integrated and the model regresses only the residual higher-order terms. The method is validated on the PhyWorld white-box physics benchmark with five tasks (uniform motion, parabola, collision, bouncing, looming), focusing on out-of-distribution scenarios to test whether the model has truly learned the underlying dynamics.

Key Results

LDR reduces the gap between in- and out-of-distribution error by over 20× compared to a video diffusion baseline, while using 26× fewer parameters and running 143× faster, and generalizes to severe distribution shifts such as predicting a blue square moving right-to-left after training only on red balls moving left-to-right.

Limitations

Evaluation is limited to a controlled white-box physics benchmark (PhyWorld) with synthetic tasks; performance on real-world, complex video dynamics is not assessed.
The abstract does not discuss limitations with stochastic or chaotic dynamics, or scalability to high-resolution natural videos.

Resources

Paper