Social-JEPA: Emergent Geometric Isomorphism
Abstract
World models compress rich sensory streams into compact latent codes that anticipate future observations. We let separate agents acquire such models from distinct viewpoints of the same environment without any parameter sharing or coordination. After training, their internal representations exhibit a striking emergent property: the two latent spaces are related by an approximate linear isometry, enabling transparent translation between them. This geometric consensus survives large viewpoint shifts and scant overlap in raw pixels. Leveraging the learned alignment, a classifier trained on one agent can be ported to the other with no additional gradient steps, while distillation-like migration accelerates later learning and markedly reduces total compute. The findings reveal that predictive learning objectives impose strong regularities on representation geometry, suggesting a lightweight path to interoperability among decentralized vision systems. The code is available at https://anonymous.4open.science/r/Social-JEPA-5C57.
Cite
@article{arxiv.2603.02263,
title = {Social-JEPA: Emergent Geometric Isomorphism},
author = {Haoran Zhang and Youjin Wang and Yi Duan and Rong Fu and Dianyu Zhao and Sicheng Fan and Shuaishuai Cao and Wentao Guo and Xiao Zhou},
journal= {arXiv preprint arXiv:2603.02263},
year = {2026}
}
Comments
This preprint is withdrawn due to significant errors in the emergent geometric isomorphism results that necessitate full rewriting, coupled with unresolved author disagreement on authorship. A corrected and revised manuscript will be released separately