Proving and improving the platonic representation hypothesis with JEPAs

The problem

Whether different models trained on separate data modalities end up with the same internal picture of the world is a central question in artificial intelligence (arxiv.org). This idea is known as the Platonic Representation Hypothesis, which suggests that various models converge to a shared statistical model of reality as they scale in size (arxiv.org). Previously, this hypothesis remained merely an observation supported by empirical trends rather than a mathematical guarantee (arxiv.org). Without a rigorous proof, researchers lacked a recipe or theoretical guidance for how to reliably achieve model alignment across modalities (arxiv.org). Instead, scaling up model sizes and dataset volumes served as the primary method to drive alignment (arxiv.org).

The idea

The authors set out to turn the hypothesis into a theorem by analyzing Joint-Embedding Predictive Architectures, known as JEPAs (arxiv.org). A JEPA is a type of self-supervised learning model trained on positive pairs of inputs (arxiv.org). The researchers prove that when separate JEPAs are trained on different modalities that observe the same underlying world, they learn the exact same representation up to a single rotation, provided their positive pairs correspond to two distinct observations of the same scene and their misalignment is otherwise bounded (arxiv.org). The authors refer to positives formed by two observations of a single world scene as good views, and contrast them with bad-view JEPAs, whose positives consist of corruptions or transformations of a single observation, such as crops of a photograph or masked text (arxiv.org). The study demonstrates that standard next-token prediction used in language models functions as a bad-view JEPA, where the positive pairs are a prefix and its continuation within one document (arxiv.org). This structural difference explains why language models align with vision only partially and mostly through massive scaling (arxiv.org).

The results

The authors validated their theory across datasets ranging from 145k Flickr captions to the texts of 940M web images (arxiv.org). They uncovered a scaling law of alignment based on the fraction of good views and the number of training samples seen, which accurately predicted 85 new training runs before they were executed (arxiv.org). Good-view JEPAs proved vastly more efficient than next-token prediction (arxiv.org). On the same text corpus, a good-view JEPA reached the same alignment with 21.6 million training samples that next-token prediction required 900 million samples to achieve, making good views 42 times more sample-efficient (arxiv.org). Furthermore, a text JEPA with 109 million parameters trained on 940 million web image captions achieved a score of .184 on the benchmark, outperforming every evaluated language model—including models up to 72 billion parameters such as Qwen2.5-72B at .154—as well as CLIP's text encoder at .124 (arxiv.org). The authors also showed that a single linear map can turn frozen DINOv2 and DINOv3 image encoders into an image captioner using a text model, achieving a CIDEr score of .838 on the COCO dataset (arxiv.org). Additionally, without using any paired data at all, their text JEPA outperformed the zero-shot accuracy of language models up to 8 billion parameters on a pair-free benchmark, scoring 64% compared to at most 56% top-5 on ImageNet-100 (arxiv.org).

Why it matters

By converting the Platonic Representation Hypothesis from an empirical observation into a mathematical theorem, this work provides a rigorous foundation for multimodal AI (arxiv.org). It demonstrates that carefully chosen positive views, rather than sheer model scale alone, act as the primary lever for building unified representations (arxiv.org). The findings indicate that current language model training practices using next-token prediction are a remarkably inefficient route to achieving platonic representations (arxiv.org). By establishing that good views require far less data and smaller model sizes to achieve superior alignment, the research offers a concrete recipe for building more efficient and unified multimodal artificial intelligence systems (arxiv.org).

Caveats

The authors note certain limitations in their theoretical framework. Their mathematical proof assumes an idealized setting with a finite world rendered injectively, population optima, and minimizers that are unique up to rotation (arxiv.org). Furthermore, the theory relies on second-order anti-collapse under a spectral gap, and full-distribution constraints such as those in LeJEPA's SIGReg target are not fully covered by the uniqueness proof (arxiv.org). The mathematical analysis also does not cover cases where embedding dimensions differ, requiring fitted linear maps with orthonormal rows in practice (arxiv.org). Finally, this work is a preprint available on arXiv (arxiv.org).