Hacker News Front Page · [Submitted on 29 Oct 2026] ·
Proving and Improving the Platonic Representation Hypothesis with JEPAs
El trabajo demuestra que JEPAs entrenadas por separado con distintas modalidades aprenden representaciones equivalentes salvo por una rotación, siempre que los pares positivos sean observaciones de una misma escena; cuando no lo son, la desalineación queda acotada. Sus experimentos muestran que estas «buenas vistas» requieren 42 veces menos muestras que la predicción del siguiente token para alcanzar la misma alineación, y que una JEPA textual de 109 millones de parámetros supera en alineación con codificadores de imágenes a los modelos de lenguaje evaluados, incluso los de hasta 72.000 millones de parámetros. Los resultados respaldan que la calidad de las vistas positivas, y no solo la escala, es clave para obtener representaciones compartidas entre modalidades.
Whether models trained on different modalities converge to the same representation of the world is central to AI. The Platonic Representation Hypothesis (PRH) conjectures that they do as they scale, but without a proof it offers neither a guarantee nor a recipe. We turn the hypothesis into a theorem. First, we prove a \emph{Platonic Representation Theorem}: Joint-Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities learn the same representation up to one rotation whenever their positive pairs are two observations of one scene--\emph{good views}--and their misalignment is bounded otherwise. Second, we show that next-token prediction is a JEPA with \emph{bad views}, parts of a single document, which explains why language models align with vision only through scale. Third, we validate the theory from 145k captions to the texts of 940M web images, and uncover a scaling law of alignment that predicts unseen runs within seed noise. Good views need 42 times fewer training samples than next-token prediction on the same corpus to reach the same alignment, and a 109M-parameter text JEPA aligns with DINOv2 and CLIP's image encoders better than every language model we evaluated, up to 72B parameters, and than CLIP's text encoder: next-token prediction is a remarkably inefficient route to platonic representations. As a showcase, one rotation turns separately trained encoders into an image captioner (CIDEr .838 on COCO). Even that rotation needs no pair: on the latest pair-free benchmark, our text JEPA beats every zero-shot accuracy reported with language models up to 8B parameters, e.g., 64\% against at most 56\% top-5 on ImageNet-100. Positive views, and not scale alone, are thus the lever to build platonic representations.