A world model that replaces attention and recurrence with reaction-diffusion PDEs. Latent states evolve through Laplacian diffusion and learned reaction terms on a 2D spatial grid, giving the model a built-in inductive bias for spatial coherence, adaptive computation depth, and O(N) scaling in the number of spatial tokens.
machine-learning deep-learning generative-model reaction-diffusion physics-simulation spatiotemporal video-prediction world-model non-transformer continuous-dynamics
-
Updated
Aug 26, 2026 - Jupyter Notebook