Runnable PyTorch implementation of the Sliced Isotropic Gaussian Regularizer (SIGReg) from the LeWorldModel paper (Maes, Le Lidec, Scieur, LeCun, Balestriero, arXiv:2603.19312, March 2026).
SIGReg is a single-term regularizer that replaces the usual six-knob stack (EMA teachers, stop-gradient, frozen pretrained features, VICReg variance/covariance terms, momentum encoders, LR warmup schedules) used to prevent representation collapse in JEPA-style self-supervised world models.
Companion post: LeWorldModel: One Regularizer Instead of Six for Pixel JEPA.