LeCun's LeVJEPA cuts video training compute by up to 20.8x
TheTuringPost · x · 2026-08-30
Yann LeCun's team introduces LeVJEPA, a method designed to make video pretraining cheaper and simpler by using a single encoder and a small projector. It employs an invariance loss on global and local video views, regularized by SIGReg to prevent representation collapse without architectural asymmetries. By randomly dropping 95% of video patches and using causal attention, LeVJEPA achieves performance matching or surpassing V-JEPA 2 while using 5.6 to 20.8x less compute at matched epochs.
Related event: LeCun Team Releases LeVJEPA: Up to 20x Cheaper Video Pretraining(6 posts)→
More from Research
- A beginner-friendly introduction to causal inference kicks off, Part 1 — Nasereliver · 2026-10-16
- FBI method trains for Pass@K with one sample per update at 12.3% of the cost — EmtiyazKhan · 2026-10-16
- Berkeley researchers sample humanoid robot policies directly from weight space with diffusion — berkeley_ai · 2026-10-16
- OneWM: unified causal training shows world and action modeling improve each other — IAAI233 · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- 5.8M-param Speech Event Detector adds laughs, breaths and背景 sounds to Whisper transcripts — Ayush Kumar Bar · 2026-10-16