Joint‑Embedding Predictive Architectures (JEPAs) are a prototypical latent world model that yields strong semantic representations. However, their dense patch embeddings expose a redundant interface to downstream predictors, which must implicitly infer spatial organization and temporal correspondence, leading to three challenges. First, object information is spread across patches, forcing each downstream task to redundantly recover object structure that could otherwise be shared. Second, the absence of explicit entity correspondence over time makes using historical transitions costly, limiting the predictor’s flexibility in leveraging context. Third, existing JEPA models typically rely on pointwise $l_1$/$l_2$ distances, which are insensitive to spatial displacement when predicted and goal supports are largely disjoint. To address these issues we introduce the Slot‑based Joint‑Embedding Predictive Architecture (Slot‑JEPA), which decomposes dense representations into stable semantic slots and time‑varying spatial masks. After a single pre‑training phase the decomposition is frozen, providing a reusable object‑structured interface for downstream tasks. Building on this, the Transition Memory Predictor (TMP) stores visual transitions as compact masks and retrieves useful context using statistics of the corresponding slots. TMP has two branches: one captures slot‑specific dynamics, the other models shared cross‑slot dynamics. We further employ Unbalanced Optimal Transport (UOT) to supervise mask prediction and to evaluate candidate plans by transporting mass between aligned predicted and goal masks, delivering a signal better aligned with spatial progress. Across seven visual‑understanding benchmarks the slot‑reconstructed features remain transferable, and Slot‑JEPA improves goal‑conditioned planning on six planar environments and four robotic‑arm tasks in the REALM‑JEPA benchmark.
Review: Slot‑JEPA’s explicit object slots and spatial masks reorganize scattered pixel information into a reusable structure, reducing the learning burden for downstream tasks, while UOT offers a geometrically meaningful supervision signal, demonstrating strong potential for multi‑task visual planning.