Paper Overview
Field: Computer Vision Authors: Haodong Li, Shaoteng Liu, Tianyu Wang Date: 2026-08-11 arXiv: 2508.03804
Motivation
The world evolves according to its dynamics—the laws of motion. Leading video diffusion models largely fit pixels without modeling how pixels transit over time, producing visually plausible frames that may not obey physical laws.
Method: Latent Dynamics Reasoning (LDR)
- Casts latent transition as explicit kinematic integration: lower-order dynamics are integrated numerically.
- The model regresses only the third- and higher-order residual that drives the rollout.
- Runs on a structured latent rather than dense convolutional features for better extrapolation.
- 20x smaller gap between in-distribution and OOD errors vs. video diffusion baselines.
- 26x fewer parameters and 143x faster inference.
- Generalizes under severe shifts: trained only on a red ball moving left-to-right, it correctly predicts the motion of a blue square moving right-to-left.
- Paper: <https://arxiv.org/abs/2508.03804>
- Project page: <https://lat-dyn-reason.github.io/>
Experiments
Following PhyWorld, LDR is validated on a controlled white-box physics benchmark covering five tasks: uniform motion, parabola, collision, bounce, and approach—with emphasis on out-of-distribution (OOD) scenarios.
Results (256^2 resolution, single-task and joint-task training)
To the authors' knowledge, this is the first video world model that extrapolates learned dynamics beyond the training distribution.