Research
Paper proves Platonic Representation Hypothesis for JEPAs and proposes text JEPA results
The paper proves a Platonic Representation Theorem for Joint-Embedding Predictive Architectures trained on multiple modalities. A 109M-parameter text JEPA reportedly aligns with DINOv2 and CLIP image encoders better than language models up to 72B parameters.
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
Proving and Improving the Platonic Representation Hypothesis with JEPAs