Agentic Learning AI Lab introduced sliced variance regularization for JEPA world models. The team says SVReg builds smooth latent maps for planning.
Stop-gradient targets and distributional regularizers already stopped collapse. The lab reports both often produced poor latent geometry. The paper is by Mengye Ren, Ying Wang, Chris Hoang, Yilun Kuang and Yann LeCun. It enters a long-running World Models debate about prediction in latent space.
SVReg imposes a variance floor on latent projections along random directions. Trajectories turn smooth and latent distances track progress in navigation and reaching.
Geometry improved. Content suffered.
SVReg can discard slowly changing, hard-to-predict factors. The example is pose of a manipulated object. The lab says stop-gradient variants tend to preserve that pose.
Slow SVReg favors slowly changing factors with a slowness term and a gated residual predictor. It keeps those factors and improves planning. It posts the highest success in hard settings across five visual control tasks. It also learns latent maps from first-person mazes and city walks. The lab tracks the release as Issue 135 in its public site repository.
Slow SVReg leads hard planning, but only on the five visual tasks tested. The filing is AGL-TR-2026-04.





