Adaptive Compute in Latent World Models: When Depth Helps, Hurts, or Doesn't Matter

arXiv:2607.10203v2 Announce Type: replace-cross Abstract: Adaptive-compute world models -- early-exit or mixture-of-depths predictors that spend variable depth per step -- assume depth buys better predictions and can be routed adaptively. In autoregressive rollouts, the first assumption requires depth's per-step precision to survive composition. We test this with a pre-registered instrument, the shallow penalty rho=err(ext{shallowest-exit rollout})/err(ext{full-depth rollout}), across nine DeepMind Control tasks under matched single-step (K=1) and multi-step (K=4) training, three seeds each. We find three regimes: on 6/9 tasks depth helps rollouts (intrinsic, rho up to 4.7imes), on 2/9 the shallow exits beat the full stack (inversion, rho down to 0.85imes), and one is flat. The robust inversion (cheetah) is not a property of the dynamics but is created by training: an ablation supervising early exits only at the first rollout step erases it (rho: 0.87o1.18, n=8, Delta=+0.31), while an intrinsic-tradeoff task is unaffected -- a double dissociation we call the routability catch-22, since the supervision that makes exits routable is what trains them to out-roll the full stack. The regime is partly predictable a priori: observation/action dimensionality and one-step model error correlate with rho at |ext{Spearman}|approx0.75 (n=9). Inside a CEM planner, rho's sign predicts whether planning benefits from depth, most sharply on the inversion task, where shallow planning beats deep. Finally, three cautions: a task's regime depends on the metric space, the rollout horizon, and the encoder. All thresholds and gates were fixed before the compute campaign, including a pre-registered negative for the hypothesis that motivated the study.

Source: arXiv cs.AI | 2026-07-15

Related