Adaptive-compute world models run fewer predictor blocks on some rollout steps than on others. This rests on two assumptions: that a deeper predictor gives a more accurate rollout, and that the steps which need depth can be identified before the predictor runs. We test both on nine DeepMind Control tasks with a twelve-block early-exit latent predictor, eight seeds per task. The first assumption fails on three tasks, where a shallower exit rolls out more accurately than the full stack on every seed, and composing predictions over ten steps moves the depth-error relation away from its one-step shape in both directions. In hindsight, choosing depth per step helps: a greedy oracle that sees the true next latent has 10-37% lower rollout error than the best mixture of fixed depths at equal average compute, and 8-28% lower than the best step-index schedule, on all nine tasks. No router we trained captures this. The difficulty router these models carry stays within 5% of fixed-depth mixtures; a router trained to predict each exit's error beats them by up to 12% but no more than a fixed schedule does; and in closed-loop planning at equal compute the routed planner matches a fixed middle depth. On cheetah, where depth 2 rolls out more accurately than the full stack (an inversion), supervising the early exits only at the first rollout step removes the inversion, while a control with the same loss weight that supervises them at one random step keeps most of it (rho 0.91 vs 0.85 standard and 1.14 first-step-only, three seeds): the inversion depends on supervising exits from composed latents. These results change with the encoder, backbone, horizon and training data. The fixed-depth campaign was pre-registered; the routing analyses, planner runs, ablation control and one-step decomposition are exploratory.

翻译：暂无翻译