Interactive World Models

Interactive world models study how to represent and predict environments as agents act, so they can support planning, policy training, and playable experiences. Common approaches include action-conditioned autoregressive video diffusion, latent dynamics models, persistent 3D scene memory, and hybrids that combine learned generation with physics or explicit world state. Evaluation spans games, robotics, driving, and interfaces, with benchmarks testing action responsiveness, spatial and physical consistency, memory on revisits, and long-horizon stability. A central tension is making rollouts fast and visually convincing while preserving causal, persistent state; models still often drift or respond inconsistently over extended interactions.