Proving the Platonic Representation Hypothesis via Joint-Embedding Predictive Architectures

The paper proves the Platonic Representation Hypothesis using Joint-Embedding Predictive Architectures (JEPAs), demonstrating that models trained on different modalities learn equivalent representations with good views. It shows JEPAs outperform next-token prediction methods, requiring 42x fewer samples for alignment and achieving state-of-the-art zero-shot performance without paired data.

Comments

Read original article