TempoRoute: Control-Relevance Routing over Temporal Skips and Recurrent Depth in World-Model Planning

Fixed-step world-model planners spend one latent transition on every imagined environment step, even when action sensitivity is concentrated around brief contacts. TempoRoute turns this mismatch into joint routing over temporal coverage and recurrent depth: at each imagined node, uncertainty growth, critic sensitivity, and latent context select an aggregate skip, a one-step standard, or a recurrent-depth burst. A physical-time actor–critic scores these operators under a common 15-step horizon and 15-query cap while receding-horizon execution retains one primitive action per observation. Across six event-heavy DMControl and ManiSkill2 tasks, TempoRoute reduces mean planner latency by 12.6% (95% CI 11.0–14.1%) and GFLOP by 27.3% (95% CI 25.9–28.6%) relative to the matched Uniform-Time planner, with 5/6 task-level quality-equivalence decisions. A histogram-matched random placement attains nearly the same cost but only 2/6 equivalence decisions, showing that where computation is placed matters beyond operator frequency. At the reported means, TempoRoute is 12.2–20.8% faster than each task's strongest-quality fixed-step comparator, with a maximum task-level mean-quality shortfall of 2.4%; complete-loop median latency falls by 10.8% on A100 and 13.9% on Jetson AGX Orin. The router also supplies its own applicability signal: smooth tasks use standard on 94.7% and 96.1% of nodes, and cue corruption at $\sigma =0.20$ contracts the three-task latency reduction from 15.4% to 1.7% [0.5, 2.9]%. These results establish node-wise time-and-depth allocation as a practical route to faster event-heavy planning without lowering the environment-facing control rate.