SpecWM: Spectral Alignment Boosts World Model Planning Success by up to 22 Points
randall_balestr · x · 2026-10-15
Researchers including William Peng, Marco Bagatella, and Randall Balestriero released a paper (arXiv:2610.15429) introducing SpecWM, a JEPA framework for world models built on spectral representation learning.
Key idea
- A target similarity kernel explicitly specifies which relationships between observations to preserve (depth, temporal distance, or other metadata), and the encoder is trained to reproduce them in embedding geometry
- Kernels range from physical-state supervision to a fully self-supervised temporal kernel requiring only trajectory ordering
- The paper characterizes global optima, conditions for linear recoverability and invariance to nuisance information, and links temporal-kernel similarity to the behavioral successor measure
Results
- Temporal spectral alignment lifts CEM planning success over LeWorldModel: OGBench Cube 52%→68%, OGBench Scene 49%→77%, CALVIN 29%→51% (avg 51% relative improvement)
- Kernel choice strongly controls which physical quantities are recoverable, showing that specifying representation geometry is a simple way to steer what latent world models preserve
Paper and blog post are publicly available.
Related event: SpecWM: Spectral Alignment Boosts World Model Planning by Up to 28 Points(2 posts)→
More from Embodied
- AlphaNIST: one RL policy solves all 20 NIST assembly tasks with 98.2% success in simulation — hbouammar · 2026-10-15
- Neural field planning for hybrid manufacturing cuts estimated build time 68.3% — jiqizhixin · 2026-10-15
- Leash Bio: DNA-encoded libraries and an RL sandbox for 10,000 drug-discovery campaigns a weekend — Latent Space · 2026-10-15
- Work photos are AI's first mile: if capture is wrong, no model can recover — arnabdotorg · 2026-10-15
- LEGO builds an omnidirectional vehicle for automated warehouses — glenbeer · 2026-10-15
- SlipSense: multimodal tactile slip detection hits 96.7% F1 at 23.1ms with zero-shot robot transfer — chris_j_paxton · 2026-10-15