The Platonic Representation Theorem explains why standard language models align with visual systems only through immense compute, revealing that next-token prediction relies on suboptimal views. By using distinct observations of identical scenes, models achieve the same level of cross-modal alignment with 42 times fewer training samples. Applying a single mathematical rotation allows these separately trained vision and text encoders to operate as functional multimodal systems, achieving strong performance on captioning tasks.
Key points
Researchers mathematically proved that separate models converge to identical representati…: Researchers mathematically proved that separate models converge to identical representations up to a rotation when trained on distinct views of the same scene.
High-quality positive training views serve as the primary driver for multimodal alignment…: High-quality positive training views serve as the primary driver for multimodal alignment rather than relying strictly on parameter scale.
The text JEPA outperformed zero-shot language models up to 8 billion parameters on ImageN…: The text JEPA outperformed zero-shot language models up to 8 billion parameters on ImageNet-100 without using paired training data.
Sources checked by our analysis pipeline: 1
Computer Science > Artificial Intelligence
[Submitted on 29 Oct 2026]
Title:Proving and Improving the Platonic Representation Hypothesis with JEPAs
View PDF HTML (experimental)Abstract:Whether models trained on different modalities converge to the same representation of the world is central to AI. The Platonic Representation Hypothesis (PRH) conjectures that they do as they scale, but without a proof it offers neither a guarantee nor a recipe. We turn the hypothesis into a theorem. First, we prove a \emph{Platonic Representation Theorem}: Joint-Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities learn the same representation up to one rotation whenever their positive pairs are two observations of one scene--\emph{good views}--and their misalignment is bounded otherwise. Second, we show that next-token prediction is a JEPA with \emph{bad views}, parts of a single document, which explains why language models align with vision only through scale. Third, we validate the theory from 145k captions to the texts of 940M web images, and uncover a scaling law of alignment that predicts unseen runs within seed noise. Good views need 42 times fewer training samples than next-token…
Commercial online marketplaces are being flooded with unauthorized cultural imitations scraped directly from community art centres. Intellectual property specialists warn that this technological exploitation threatens the livelihood of First Nations creators while stripping their traditional knowledge of authenticity. In response, legal experts are urging the federal government to establish enforceable safeguards against uncompensated data scraping.
Researchers tested autonomous multi-agent dynamics by deploying model swarms into virtual environments using existing scripting frameworks like Mindcraft and Mineflayer. The trials revealed that collaborative tasks can rapidly collapse into crippling procedural delays under minor environmental pressures, even as similar architectures excel in head-to-head competitive settings. These open-sourced findings offer critical real-world benchmarks for understanding the boundaries and operational risks of large-scale agen…
Recent discussions have centered on the significant potential for advanced computational technology to preserve human civilization. Analysts evaluating these scenarios have focused heavily on the potential positive outcomes rather than purely existential risks. This perspective suggests that technological advancement could serve as a decisive protective force.