arXiv:2608.04135 2026 Architecture 1 ideas extracted · analyzed Aug 31, 2026
The paper turns support-constrained coupling into a graph-theoretic feasibility problem: a binary support pattern can carry prescribed marginals exactly when every subset satisfies a Hall-type neighborhood inequality. This gives a concrete preconditioner for sparse attention, mixture-of-experts routing, or masked transport layers: before training, certify that the mask admits a nondegenerate flow rather than discovering infeasibility through dead tokens or unstable normalization. The most practical transfer is a Hall-violation diagnostic and mask-repair procedure followed by Sinkhorn or constrained flow normalization; it is more useful as a structural guarantee than as a new optimizer.
Use the support graph of a sparse attention or routing mask as a feasibility object before applying normalization. Reject or minimally repair masks whose neighborhoods violate Hall's condition, ensuring that every selected query/key subset has enough reachable mass to support the desired marginals.