Researchers Prove Platonic Representation Hypothesis

Researchers prove the Platonic Representation Hypothesis by showing Joint-Embedding Predictive Architectures learn identical representations up to rotation. Validated across 145k captions and 940M web images, the theory reveals good views need 42 times fewer samples than next-token prediction to match alignment. A 109M-parameter text JEPA beats models up to 72B parameters, achieving 64% accuracy against 56% top-5 on ImageNet-100.