Tag
Introduces Latent Dynamics Reasoning (LDR), a video world model that integrates kinematic dynamics in a structured latent space, enabling extrapolation of learned dynamics far beyond training distributions while using far fewer parameters and running much faster than video diffusion baselines.