LDR Video World Model Extrapolates Dynamics Beyond Training Distribution

Scobleizer · x · 2026-08-25

Latent Dynamics Reasoning (LDR) is a new video world model designed to understand "how the world evolves." Instead of predicting future frames directly, it maps past frames to structured latent states and rolls them forward via kinematic integration, learning only the high-order motion residuals. The authors claim LDR is the first video world model to extrapolate learned dynamics beyond its training distribution. Paper, code, models, and datasets are now publicly available.

Related event: Adobe's LDR World Model Extrapolates Beyond Training Distribution(3 posts)→

Original post →

More from Research

Research channel →