Paper Overview
- Research Area: Computer Vision (CV)
- Authors: Haodong Li, Shaoteng Liu, Tianyu Wang
- Release Date: 2026-08-11
- arXiv: 2508.03804
- Project Page: https://lat-dyn-reason.github.io/
- Lower-order dynamics are integrated numerically.
- The model regresses only the third- and higher-order residual that drives the rollout.
- LDR achieves >20x smaller gap between in-distribution and OOD error compared to video diffusion baselines, in both single-task and joint-task training at 256² resolution.
- Uses 26x fewer parameters than baseline diffusion models.
- Runs 143x faster.
- Demonstrates strong generalization under severe distribution shifts—for example, trained only on red balls moving left to right, LDR correctly predicts the motion of a blue square moving right to left.
Summary
The world evolves according to its dynamics—the laws of motion. However, leading video diffusion models largely fit pixels without explicitly modeling how those pixels change over time. As a result, they generate visually plausible frames but may fail to obey physical laws.
To capture dynamics purely from pixels, the authors introduce Latent Dynamics Reasoning (LDR). LDR casts the latent transition as an explicit kinematic integration:
For better extrapolation, LDR operates on a structured latent instead of dense convolutional features.
Evaluation
Following the PhyWorld benchmark, LDR is validated on a controlled white-box physics benchmark spanning five tasks:
1. Uniform motion 2. Projectile motion 3. Collision 4. Bouncing 5. Approach
The focus is on out-of-distribution (OOD) scenarios, testing whether the model has genuinely learned underlying dynamics.
Key Results
Significance
According to the authors, LDR is the first video world model that extrapolates learned dynamics beyond the training distribution, opening a path toward world models grounded in physics rather than pixel statistics alone.
---
*Auto-collected from arXiv on 2026-08-12*