arXiv 2511.08544

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

By Randall Balestriero and Yann LeCun

Published 2025-11-11

Mindmap

Browse the paper's core ideas, clusters, and relationships in a structured outline.

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