New Relation architecture replaces Attention with explicit Self/Exchange relation decomposition

青稞AI · wechat · 2026-09-04

The author documents two papers: MemGLU, testing whether SwiGLU's open positive tail is necessary via a closed-tail gating function (which drew a code request from SELU author Günter Klambauer within a day of arXiv posting), and the main work, "Ask Self, Ask Others: Relation Is All You Need."

Key idea: split Attention's fused relation-formation and information-routing into separate steps. Token relations are decomposed into diagonal Self and off-diagonal Exchange roles, projected by a pair of nonlinear functions chosen after 23 ablations, with a log correction term fixing Exchange's natural growth with sequence length. The multi-head version uses alternating Givens rotations to decouple relations from information across heads. Because the relation matrix is explicit, the author derives a FlashRelation blockwise computation path needing only running statistics — opening a route toward Linear variants.

Original post →

More from Research

Research channel →