Latent-Space Predictive Learning
Summary
Latent-space predictive learning trains models to predict future representations, not only future raw observations.
What The Wiki Currently Believes
- Eidos uses latent-space predictive learning for time-series forecasting robustness.
- LeWorldModel predicts future latent states conditioned on actions for control.
- NEPA predicts future visual patch embeddings.
- Reconstruction or Semantics? evaluates which latent spaces make robotic diffusion world models useful.
Evidence
The corpus repeatedly treats latent prediction as a way to suppress irrelevant surface noise while retaining task-relevant dynamics.
Open Questions
- Which latent targets are most stable: learned online targets, pretrained semantic encoders, or distribution-regularized embeddings?
- How should latent objectives stay grounded enough for high-fidelity generation?