World Modeling through Spectral Alignment
We introduce SpecWM, a JEPA framework that explicitly specifies desired relationships between observations through a target similarity kernel and trains the encoder to reproduce those relationships in its embedding geometry.
ProofPaper ↗
Key points
- How do we tell a world model what information to preserve?
- We study this question through the lens of spectral representation learning.
- More importantly, temporal spectral alignment improves CEM planning success over LeWorldModel from 52% to 68% on OGBench Cube, 49% to 77% on OGBench Scene, and 29% to 51% on CALVIN.
- These results suggest that explicitly specifying representation geometry provides a simple way to steer what latent world models preserve for control.
Sources (1)
- [1]World Modeling through Spectral AlignmentarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 08:05 AM
We introduce SpecWM, a JEPA framework that explicitly specifies desired relationships between observations through a target similarity kernel and trains the encoder to reproduce those relationships in its embedding geometry.
How do we tell a world model what information to preserve?
Extractive summary: sentences quoted from the sources.
Before this
- Oct 11, 2026LeHami: Symplectic Latent Dynamics for World Models
- Oct 10, 2026MORE: Learning Motion-Oriented Registers for Driving World Models and Planning
- Oct 9, 2026Sparse-TD-JEPA: Improving Zero-Shot Reinforcement Learning With Sparse Latent Task Representations
- Oct 9, 2026GLIMPSE: Global-Local Indexing with Multi-Precision Spectral Encoding for Explorable Video World Models
- Oct 9, 2026RSI-Anything: Unlocking Recursive Self-Improvement with World Models
- Oct 9, 2026Latent World Action Model for GUI Agents