The Geometry of Convergence
Balestriero and LeCun formally prove the Platonic Representation Hypothesis for JEPAs: models trained separately on any modality learn the same representation up to rotation, provided they use good positive pairs. The result reframes next-token prediction as a JEPA with structurally bad views, explains the slow alignment of language models with vision, and shows a 109M text JEPA outperforms every LLM tested up to 72B parameters on cross-modal tasks.
Read more →