arXiv:2608.17130 2026 Architecture 1 ideas extracted · analyzed Sep 1, 2026
The paper provides a discrete-convex projection mechanism for converting arbitrary polymatroid bases into bases obeying coordinatewise capacity constraints, while preserving the exchange structure that makes local repairs possible. This is transferable to constrained mixture-of-experts routing: expert loads, group capacities, or hierarchical token-allocation constraints can be represented as integer base vectors, and cage retraction can replace ad hoc capacity dropping and rerouting. The key engineering asset is not merely projection, but the symmetric exchange axiom, which yields short local swap sequences between feasible allocations. A practical first use is a hard-capacity router whose logits produce an unconstrained assignment and whose assignments are repaired by exchange moves before dispatch.
Replace overflow dropping in a mixture-of-experts layer with a discrete-convex load repair procedure. The router first chooses experts from neural logits, then applies capacity-aware exchange moves that preserve the total number of dispatched tokens and monotonically improve the routing objective whenever a feasible swap exists.