Looped World Models

The authors introduce Looped World Models (LoopWM), a novel architecture that utilizes iterative refinement of latent environment states through a parameter-shared transformer block, achieving up to 100x parameter efficiency compared to traditional world models. This approach allows for adaptive computation that scales the model's depth according to the complexity of each prediction step, establishing iterative latent depth as a new scaling axis for world simulation. This advancement is significant for practitioners as it enhances long-horizon simulation capabilities while reducing deployment costs and mitigating compounding errors.