Proving and Improving the Platonic Representation Hypothesis with JEPAs

Researchers prove that Joint-Embedding Predictive Architectures (JEPAs) trained on different modalities converge to the same representation of the world, converting the Platonic Representation Hypothesis into a formal theorem. They demonstrate that using "good views" (multiple observations of the same scene) is far more efficient than next-token prediction for achieving cross-modal alignment, with a 109M-parameter text JEPA outperforming language models up to 72B parameters. The findings show that the quality of training data pairs, rather than scale alone, is the key factor in building unified representations across modalities.