LeJEPA latents recover world state up to rotation, LeCun team proves
New proof from LeCun and Balestriero establishes that LeJEPA's Gaussian-regularized embeddings linearly reconstruct ground-truth latent variables, enabling direct planning in learned space.
LeJEPA—the Joint-Embedding Predictive Architecture from Yann LeCun's lab—now has a formal identifiability proof. A paper by David Klindt, LeCun, and Randall Balestriero demonstrates that LeJEPA's combination of alignment loss and isotropic Gaussian regularization linearly recovers the true latent variables of a generative world model from nonlinear observations, accurate up to an orthogonal rotation. In a broad class of stationary environments with additive noise, the Gaussian prior is the only distribution guaranteeing such precise reconstruction—a result that inverts decades of nonlinear ICA folklore, which held Gaussians as the one case where source separation fails entirely.
The practical payoff: learned representations are stable enough to run standard control algorithms—Linear-Quadratic Regulator, for example—directly on top of a pretrained encoder, with no manual tuning. The authors validated the theory on toy environments and showed that BatchNorm is critical; without it, over a third of training runs collapse. The proof covers stationary Markov Decision Processes with additive Gaussian noise and establishes that orthogonal identifiability is theoretically sufficient for optimal planning in latent space. Code is available on GitHub; the arXiv preprint (2605.26379) was posted June 4, 2026.
