New CTWM paper shapes world model latent spaces so distances aid planning
hisspikeness · x · 2026-10-05
A new paper led by Michael Hauri & Peter Buttaroni, "Learning Commute-Time-Preserving World Models for Planning" (arXiv:2610.01373), tackles what "distance" should mean in an agent's self-learned latent map.
- Key idea: latent planners pick actions reducing goal distance, but the right notion of closeness is commute-time distance — how long a random walker takes to reach a goal and return, capturing bottlenecks and shortcuts.
- Problem: graph Laplacian spectral embeddings have this property under eigenvalue-dependent scaling, but are intractable in large continuous environments; common self-supervised methods that encourage isotropic representations degrade that scaling.
- Method: CTWM combines a latent displacement predictor with a log-determinant regularizer that provably recovers the correctly scaled Laplacian representation under reversible deterministic dynamics.
- Results: CTWM matches or beats task-agnostic baseline LeWM on several continuous goal-reaching benchmarks with half the parameters.
The thread frames the work via Tolman's latent learning: rats build maps of mazes before any reward exists.
More from Research
- Schmidhuber: Linearized self-attention Transformers date back to his 1992 NECO paper — SchmidhuberAI · 2026-10-05
- ML Conference AC warns: papers that are 'incomprehensible' will be desk rejected — tyrell_turing · 2026-10-05
- Meta's ProWAM hits 70% zero-shot real-world robot success with sparse visual sub-goals — meta · 2026-10-05
- Interpretability Researcher: Don't Let Unexplained Model Behavior Block You — repligate · 2026-10-05
- Yandex Music replaced 15+ candidate generators with one transformer, +6.3% listening time — SettingAccording8986 · 2026-10-05
- Cohere Labs lands multiple papers and a workshop talk at COLM — Cohere_Labs · 2026-10-05