A Generalization Theory for JEPA-Based World Models

A new theoretical framework for Joint Embedding Predictive Architectures (JEPAs) has been submitted to the arXiv preprint server, offering the first generalization theory for this class of world models, according to a paper dated 25 Jun 2026 [1]. JEPA-based world models operate by learning predictive dynamics in a latent space, an approach that differs from models that generate future observations at the input level [1]. The theoretical understanding of these models had remained limited despite their empirical success [1]. The new paper formulates JEPA pretraining as a conditional spectral graph learning problem and demonstrates that the JEPA objective is equivalent to a low-rank factorization of an action-conditioned co-occurrence matrix [1]. Building on this characterization, the authors establish a connection between JEPA pretraining error and downstream planning regret, which yields a finite-sample generalization bound [1]. The analysis reveals an inherent trade-off between approximation and sample errors with respect to the latent dimension, providing insight into the advantages and limitations of latent predictive models compared with input-level predictive approaches [1]. The work is situated within the broader field of self-supervised learning (SSL), a paradigm where models generate their own supervisory signals from the structure of the data itself rather than relying on externally-provided labels [2]. SSL tasks are designed to require the capture of essential features or relationships in the data, often through augmentation techniques such as introducing noise, cropping, or rotation [2]. The paper was posted on arXiv, an open-access repository for electronic preprints that has served the physics, mathematics, and computer science communities since its founding on August 14, 1991 [6]. As of November 2024, the repository receives about 24,000 new articles per month and had surpassed two million total articles by the end of 2021 [6]. Papers on arXiv are moderated but not peer-reviewed before posting [6]. The platform also supports community-developed tools through its arXivLabs framework, which allows third-party collaborators to build features such as citation explorers and recommender systems directly on article pages [5].