Joint-Embedding Predictive Architectures (JEPAs) are a common design for latent world models, prized for producing strong semantic representations. But the new preprint argues that their dense patch representations expose a redundant interface to any downstream predictor that must work from them.
The proposed remedy, Slot-JEPA, swaps those dense patches for object-structured prediction interfaces built around slots. This is intended to give predictors a more compact and semantically meaningful input, rather than forcing them to sift through redundant low-level details.
Because the abstract is truncated, the paper's experimental results and implementation specifics are not available from this source. This is a single preprint, so there are no contrasting findings to compare; the claims here are about the design rationale and the proposed architecture itself.