arXiv:2607.19940 2026 Architecture 1 ideas extracted · analyzed Aug 30, 2026
The paper identifies a large Boolean family inside the partition lattice by representing partitions as connected-component decompositions of a tree. Every subset of tree edges produces a globally valid partition, so grouping decisions are independent edge coordinates rather than unconstrained pairwise merges that may create inconsistent or cyclic structures. This suggests a neural token-routing module with learnable edge gates on a fixed tree, providing hierarchical grouping with controllable computation and an explicit validity guarantee. The first implementation should use soft gates during training and connected components with union-find at inference.
Replace unconstrained pairwise token grouping with a tree whose edges carry independent merge or cut variables. The connected components of the retained edges define a valid partition at every forward pass, while learned edge gates control the amount of token aggregation. A coarse component-level computation can then replace part of dense attention.