LeCun and Balestriero prove separately trained models converge, and a 109-million-parameter text model beats 72-billion-parameter ones at it
Randall Balestriero and Yann LeCun posted a paper on October 29th proving that joint-embedding predictive architectures (JEPAs) trained separately on different kinds of data learn the same representation up to a single rotation, provided each training pair is two views of the same scene. They argue next-token prediction is the same setup with poor pairs, which is why language models only line up with vision models at scale. Good pairs reached the same with 42 times fewer samples.
A 109-million-parameter text aligned with image encoders better than every language model tested, up to 72 billion , and one rotation turned two independently trained encoders into an image captioner. The results are on alignment tests the authors chose, and the paper is not peer reviewed. SourcesA
Sources
This story ran in the Morning Brief of November 2nd 2026.