Does Depth Survive Composition, and Can a Router Use It? Adaptive Compute in Latent World Models
The study investigates adaptive compute world models that use variable predictor depth per step to save computation. It found an 'inversion' phenomenon where shallow rollouts outperform full-depth ones, caused by training that makes early exits routable. This suggests the benefit of more compute is configuration-dependent and not an intrinsic property of task dynamics.