This is a Plain English Papers summary of a research paper called How JEPA Positive Pairs Help Prove and Improve Cross-Modal Alignment. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.
Shared positive views turn PRH into a theorem
A 109M-parameter text JEPA trained on captions of web images reaches.184 alignment with DINOv2, exceeding every language model tested, including models up to 72B parameters. The result supports the paper’s central claim: separately trained encoders can learn a shared representation when their positive pairs describe the same world states, rather than merely sharing a training scale.
The Platonic Representation Hypothesis, or PRH, proposes that models trained on different modalities converge toward a common representation of the world. Balestriero and LeCun formalize this for JEPAs: if modalities render the same finite world and use the same distribution of positive pairs, their optimal embeddings agree up to an orthogonal rotation, assuming the shared objective has minimizers unique up to rotation.
The theorem is deliberately idealized. It assumes injective renderings, population-level optimization, and shared positive-pair distributions; it does not establish that arbitrary practical encoders will align. For whitened JEPAs, the authors show uniqueness under a spectral gap, with the learned embedding given by the top eigenvectors of the graph induced by positive pairs.
