社会-JEPA：新兴几何同构

This research presents a novel world model that compresses rich sensory streams into compact latent codes to anticipate future observations. Multiple agents independently acquire such models from distinct viewpoints of the same environment without parameter sharing or coordination. After training, the internal representations of the agents exhibit a striking emergent property: the two latent spaces are related by an approximate linear isometry, enabling transparent translation between them. This geometric consensus persists despite large viewpoint shifts and minimal overlap in raw pixels. Leveraging the learned alignment, a classifier trained on one agent can be seamlessly transferred to another without additional gradient steps, while distillation-like migration accelerates subsequent learning and significantly reduces total learning time.