arXiv:2608.13727 2026 Architecture 1 ideas extracted · analyzed Sep 1, 2026
The paper's useful transferable asset is the Jucys–Murphy construction: sums of transpositions that commute, admit simultaneous diagonalization, and expose a structured spectral basis for permutation representations. This suggests replacing an unstructured token-mixing block with a learnable spectral filter over commuting swap operators, giving a controllable hierarchy between full permutation symmetry and ordered interactions. The hitting-time result itself is less directly useful for neural networks, since coupon-collector-style star shuffling is not an efficient alternative to standard random permutation generation.
Add a structured token-mixing layer based on commuting sums of swap operators rather than unconstrained pairwise attention. The layer learns a low-degree spectral filter in the Jucys–Murphy operators, allowing it to represent hierarchical interactions while retaining an explicit algebraic inductive bias.