CTWM: Commute-Time-Preserving World Model Matches Strong Baseline with Half the Parameters

Michael Hauri, Peter Buttaroni, Friedemann Zenke, and colleagues published a paper on arXiv, "Learning Commute-Time-Preserving World Models for Planning" (2610.01373), proposing the Commute-Time World Model (CTWM), whose core goal is to make distances in a world model's latent space genuinely meaningful for planning.

Confirmed

Across six pixel-based goal-reaching tasks (covering navigation and manipulation), CTWM matches or exceeds the strong task-agnostic baseline LeWM with only half the parameters (9M vs 18M); on PointMaze, its latent space was highlighted as better reflecting the maze's true structure.
Theoretical basis: if the latent space aligns with the environment's graph Laplacian eigenvectors and is scaled correctly, straight-line distances in the latent space equal commute-time distances — the time a random walker needs to reach a goal and return. This metric captures bottlenecks and shortcuts, making it better suited for planning than generic "latent distance."
The key problem is that most self-supervised learning pushes representations toward a uniform distribution, which precisely destroys the properties needed for commute-time scaling. The authors combine residual prediction with a log-determinant regularizer and prove this recovers correctly scaled representations.

Why it matters

The results show that with the right metric (aligning with the graph Laplacian and scaling correctly), latent distances yield planning signals with genuine geometric and graph-theoretic meaning, giving latent-space planning in world models a firmer theoretical foundation.
The log-det regularization fix for self-supervised representation scaling could also benefit other methods that rely on latent-space distances; meanwhile, halving the parameters makes deployment friendlier.

2026-10-05 ~ 2026-10-05 · 6 related posts

Primary sources

CTWM matches or beats LeWM on six pixel-based tasks with half the parameters — hisspikeness · 2026-10-05 15:36 (UTC+08:00)
New CTWM paper shapes world model latent spaces so distances aid planning — hisspikeness · 2026-10-05 15:32 (UTC+08:00)
CTWM world models preserve commute-time structure for planning with half the parameters — hisspikeness · 2026-10-05 15:37 (UTC+08:00)

[source] New CTWM paper shapes world model latent spaces so distances aid planning — hisspikeness · 2026-10-05
Aligning latent space with graph Laplacian eigenvectors yields commute-time distance — hisspikeness · 2026-10-05
Residual prediction plus a log-determinant regularizer recovers correctly scaled representations — hisspikeness · 2026-10-05
[source] CTWM matches or beats LeWM on six pixel-based tasks with half the parameters — hisspikeness · 2026-10-05
CTWM beats 18M-param LeWM on six goal-reaching tasks with half the parameters (9M) — hisspikeness · 2026-10-05

1 near-duplicate retellings: hisspikeness