LeCun and Balestriero prove Platonic Representation Theorem for JEPAs

Two neural networks trained on different modalities learn the same representation, up to one rotation. That is the formal theorem Randall Balestriero and Yann LeCun claim to have proven for Joint-Embedding Predictive Architectures, in a paper posted to arXiv titled 'Proving and Improving the Platonic Representation Hypothesis with JEPAs.'

The conditions are specific. The architectures' positive pairs must be two observations of one scene, what the authors call 'good views,' with misalignment bounded otherwise. Where those conditions hold, the paper argues the separate training problems are the same problem written in different coordinate systems, with solutions coinciding up to a rotation.

The empirical companion is louder. Their 109-million-parameter text JEPA 'outperforms language models up to 72B parameters' in alignment metrics, the authors report, and 'next-token prediction is a JEPA with bad views.' Good views, they argue, need 42 times fewer training samples to reach equivalent alignment. A single rotation, in their reported setup, lets one trained encoder serve as an image captioner, hitting CIDEr .838 on COCO.

The through-line is one sentence: 'positive views, and not scale alone, are thus the lever to build platonic representations.'

The abstract names figures without the specific benchmark suite behind 'alignment metrics,' the full training recipe, or how 'good views' are constructed outside paired-observation settings.