Slot-based object-centric learning assigns each object in a scene to a discrete “slot” — a learned representation that captures one entity. The slots compete for image regions through an attention mechanism: each pixel is claimed by whichever slot attends to it most strongly. When the number of slots exceeds the number of objects, multiple slots claim overlapping regions of the same entity. Two slots fight over one object. Neither wins cleanly. The segmentation degrades.
Chatzisavvas et al. (arXiv:2603.11246, 2026) observe that the competition itself identifies the problem. Slots that overlap heavily — measured by Soft-IoU between their attention masks — are redundant representations of the same entity. The fix is to merge them: combine overlapping slots into a single representation using barycentric updates that preserve gradient flow. The merging is a fixed policy, requires no additional learned parameters, and operates as a drop-in addition to existing slot-attention architectures.
The structural point: the failure mode — slots competing for the same region — produces the exact signal needed to correct it. High overlap between two slots means they are representing the same object, and merging them is the appropriate response. A system that treated overlapping slots as a training failure to be minimized through loss design would fight the symptom. Merging treats the overlap as information: these two things are actually one thing, and the competition between them is evidence.
The method reverses the usual approach to the “too many slots” problem. Most solutions try to prevent over-allocation (learn the right number of slots) or penalize redundancy (add a diversity loss). Slot merging accepts that redundancy will occur and uses it constructively — the redundancy is how the system discovers that two representations should be one.