Morning Edition · Sunday, September 13, 2026Published at 2:28 AM EDT · New York

A Collapse-Free Video Encoder Matches V-JEPA 2 With Up to 20.8 Times Less Pretraining Compute

LeVJEPA drops 95 percent of patch tokens and replaces the usual self-supervised heuristics with a single regularizer that provably rules out representation collapse, and at equal compute it leads the strongest video baseline by 7.6 points on ImageNet-1K.

A paper from Lukas Kuhn, Lucas Maes, Giuseppe Serra, Quentin Le Lidec, Yann LeCun, chief AI scientist at Meta, Randall Balestriero and Florian Buettner introduces LeVJEPA, the first video encoder trained under the LeJEPA objective. The meth…