arXiv 2511.08544

LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics

By Randall Balestriero and Yann LeCun

Published 2025-11-11

Citation lineage

Review the prior work and downstream research connected to this paper.

Learning manipulable representations of the world and its dynamics is central to AI. Joint-Embedding Predictive Architectures (JEPAs) offer a promising blueprint, but lack of practical guidance and theory has led to ad-hoc R&D. We present a comprehensive theory of JEPAs and instantiate it in LeJEPA , a lean, scalable, and theoretically grounded training objective. First, we identify the isotropic Gaussian as the opt…

View the original paper on arXiv