Proving and Improving the Platonic Representation Hypothesis with JEPAs

Researchers Randall Balestriero and Yann LeCun mathematically prove the Platonic Representation Hypothesis, demonstrating that Joint-Embedding Predictive Architectures trained on separate modalities learn identical representations when using positive observation pairs of the same scene. Their analysis reveals that next-token prediction relies on bad views and is an inefficient route to representation alignment compared to positive views. The authors validate their theory with extensive experiments, showing that a text JEPA outperforms much larger language models on zero-shot image classification tasks. Open article →