Proof: Shared Observations Beat Scale for Vision-Language Alignment

The Platonic Representation Hypothesis—the notion that AI models trained on different modalities converge to identical representations as they scale—has been transformed from conjecture into proven mathematical fact. Researchers demonstrated that Joint-Embedding Predictive Architectures (JEPAs) learn identical representations when trained separately on any number of modalities, provided they observe “good views” (multiple observations of the same scene). This proof establishes not just that convergence happens, but the precise conditions enabling it.

The work exposes a fundamental inefficiency in how language models are trained. Next-token prediction, the standard approach, functions as a JEPA using “bad views”—fragmentary document excerpts rather than coherent paired observations. This explains why language models require massive scale to achieve vision alignment. Empirical testing across 940 million web images and 145,000 captions shows training with good views requires 42 times fewer samples than next-token prediction to reach equivalent alignment.

The practical implications are substantial: a 109-million-parameter text JEPA outperforms language models with 72 billion parameters at vision-language tasks, and exceeds zero-shot performance of much larger language models. The findings demonstrate that how models are trained matters as much as their size, shifting focus from scale alone to the structure and nature of training observations.