When Does Depth Survive Composition? Compute--Quality Regimes in Latent World Models
2026-07-11 • Machine Learning
Machine LearningArtificial Intelligence
AI summaryⓘ
The authors studied how models that change their computation depth during predictions perform on several control tasks. They tested whether using more computation depth improves predictions over multiple steps and found three patterns: on some tasks, deeper computation helps; on others, shallower computation is better; and on some, there's no difference. They discovered that how these models are trained affects whether deeper predictions are actually helpful or not. Additionally, they found some task features can predict these outcomes, and the benefits depend on factors like how the task is measured and the model's input representation.
adaptive computationworld modelsearly exitautoregressive rolloutDeepMind Control Suitemodel precisionCEM plannertraining supervisionSpearman correlationrollout horizon
Authors
Achyuthan Sivasankar
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 $ρ=\mathrm{err}(\text{shallowest-exit rollout})/\mathrm{err}(\text{full-depth rollout})$, 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, $ρ$ up to $4.7\times$), on 2/9 the shallow exits beat the full stack (inversion, $ρ$ down to $0.85\times$), 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 ($ρ: 0.87\to1.18$, $n=8$, $Δ=+0.31$), 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 $ρ$ at $|\text{Spearman}|\approx0.75$ ($n=9$). Inside a CEM planner, $ρ$'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.