It's the Model, Not the Macro-Actions: What Hierarchy Buys in Latent-Space Planning

A hierarchical latent world model raises two abstraction levels at once: it rolls plans through a model that predicts ten steps at a time, and it searches over learned temporally extended macro-actions. The pair beats flat planners on long-horizon tasks, and the standard explanation credits the action abstraction. Using a released model pair whose flat and hierarchical checkpoints have bitwise-identical level-1 weights — so arms differ only in the planner and in which of the two predictors rolls the plan — we find that, on this pair, the credit belongs elsewhere. We first characterize what the learned macro-action space contains: a $\sim 3.5$-dimensional code whose variance the two lowest temporal modes of a ten-step action chunk reach to $98\%$ and plain constant-hold chunking to $65.5\%$. We then run the same hand-designed action classes through each of the two models. Over the fine model, every restriction scores below the unrestricted planner in the same harness — $5/20$, $2/20$, $0/20$, $2/20$ across classes spanning $66\%$ to $99\%$ coverage, against $9/20$ (three of the four gaps significant). Over the coarse model, the crudest class — two numbers per macro-action — reaches $19/20$, matching the full released hierarchy, with 14 discordant trials to 0 against the identical class over the fine model ($p=1.2\times 10^{-4}$); searching the entire learned $8$-D space does no better ($17/20$). Model-side measurements taken with no planner in the loop suggest why: at long lookahead the fine model's cost stops ordering plans the way the environment does (latent cost, $\rho =0.20\to 0.05$ from $100$ to $250$ steps; position cost, $-0.07$ at $500$), while the coarse model's does not, and the gap is significant at $500$ steps ($+0.15$, $p=0.0012$, 17/20 layouts, $K=256$ plans). At $250$ steps an earlier $K=64$ estimate of that gap ($+0.14$, $p=0.007$) did not survive a $4\times$ larger plan sample ($+0.05$, $p=0.26$); we report it as within noise. We also report what failed: truncating the flat planner's lookahead to where its model is still trustworthy does not help, monotonically, across a $10\times$ range ($1/20$, $5/20$, $10/20$, $12/20$), because the task's own objective is uninformative at short horizons ($\kappa =0.007$ at 50 steps). Of this paper's preregistered predictions three were confirmed, three refuted and two null: the confirmations are two model-side measurements and the prediction that constant hold over the coarse model reaches $\ge 17/20$; every refutation is a prediction about the flat planner at its own timescale.