SURVEYOR: Certified Speculative Plan Consumption for Latent World-Model Planning

Hierarchical latent planners extend the reach of world-model-based control by decomposing distant goals into subgoal sequences. However, regenerating these sequences at every replanning step incurs computation even when existing drafts remain consistent with execution. We introduce SURVEYOR , a training-free framework that adaptively reuses latent plans and determines when hierarchical drafting is needed. At each replanning boundary, SURVEYOR compares the achieved latent state with the waypoint just pursued. It reuses the remaining draft when their discrepancy falls within a calibrated tolerance and regenerates the sequence from the achieved state when the check fails or the block is exhausted. The tolerance is calibrated offline using observation pairs identified as task-equivalent under the benchmark's success criterion. Using the same tolerance, an arbiter evaluates the executor's predicted goal reachability to bypass drafting initially or retire it during execution. Across mutiple environment--horizon settings on PushT, Reacher, and OGBench-Cube, SURVEYOR matches or outperforms the strongest flat baseline while substantially reducing high-level planner calls relative to every-step drafting. It also transfers to an amortized executor without additional training, increasing success from $65.6\%$ to $93.0\%$ in that executor's weakest long-horizon setting while reducing episode completion time.