Generalizations of nets and Latin squares
arXiv:2607.10890
2026
Architecture
2 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The paper develops reticulations and cooperative systems: several line partitions of a point grid whose cross-type intersections are unique, represented by row-Latin and column-Latin matrices with cross-family orthogonality. The transferable asset is deterministic collision control: Latin constraints provide exact per-row and per-column balance, while orthogonality ensures that pairs of labels identify grid positions uniquely. This can replace random hashing in sparse attention or mixture-of-experts routing with structured assignments that have provable coverage and load properties. The most practical first test is a Latin-square family used as multiple sparse attention masks or expert-routing hashes, compared against random hashing at identical density.
Ideas from this paper
✗ Failed on benchmark
2026
Partition tokens into a rectangular grid and use Latin-square labels to define several sparse attention heads. Each head attends only to tokens sharing a row, column, or Latin label, while orthogonality guarantees that every pair of labels occurs at most once, reducing systematic blind spots and repeated collisions. The resulting masks are deterministic, reusable across examples, and can be generated without learned routing scores.
Useful7/10
Difficulty5/10
Novelty7/10
Unverified
2026
Use mutually orthogonal Latin labels as deterministic expert assignments for token batches. Each routing family is individually balanced, and pairs of families avoid repeated co-assignment patterns, enabling multiple routing rounds or auxiliary experts without the severe load collisions caused by independent random hashing.
Useful6/10
Difficulty5/10
Novelty6/10