arXiv:2610.39302 Proves Platonic Representations via JEPAs Require Bounded Misalignment on Good Views

The paper converts the Platonic Representation Hypothesis into a theorem for Joint-Embedding Predictive Architectures. It demonstrates that next-token prediction functions as a JEPA with bad views from document fragments, requiring scale to achieve cross-modal alignment. Experiments scale from 145000 captions through 940 million web images and establish a scaling law of alignment that predicts held-out runs inside seed variance.

Data show a 109 million parameter text JEPA surpasses every evaluated language model up to 72 billion parameters and CLIP text encoder in alignment with DINOv2 and CLIP image encoders. Good views reach equivalent alignment with 42 times fewer samples than next-token prediction on identical corpora. One learned rotation converts independently trained encoders into an image captioner scoring CIDEr 0.838 on COCO without paired supervision.

The result isolates positive views as the operative variable rather than parameter count. Prior alignment studies that relied on next-token objectives therefore measured an inefficient path. Operational pipelines can now substitute JEPA objectives with curated multi-view data to reduce compute while meeting alignment thresholds.

Subsequent work will test whether the rotation operator generalizes to additional modalities and whether the misalignment bound can be relaxed without loss of the uniqueness guarantee.