Yann LeCun JEPA paper: 109M text model out-aligns LLMs up to 72B
Why it matters
Randall Balestriero and Yann LeCun turn the Platonic Representation Hypothesis into a theorem: JEPAs trained separately on different modalities learn the same representation up to one rotation when positive pairs are two observations of one scene (good views). They argue next-token prediction is a JEPA with bad views (parts of one document), explaining why language models align with vision only through scale. On the same corpus, good views needed 42x fewer training samples for the same alignment, and a 109M-parameter text JEPA aligned with DINOv2 and CLIP image encoders better than every language model they evaluated, up to 72B parameters. The result is about representation alignment, not overall model ability.
Read the source: arXiv (opens in a new tab)
All hits from 1 Nov 2026More ResearchGet the daily hits by email