Scaling Bidirectional Spans and Span Violations in Attention Mechanism

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Kim, Jongwook, Yun, Sangheon, Yoon, Sukjin
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914201715343360
author Kim, Jongwook
Yun, Sangheon
Yoon, Sukjin
author_facet Kim, Jongwook
Yun, Sangheon
Yoon, Sukjin
contents The canonical $O(N^2)$ Transformer remains the empirical performance frontier in sequence modeling, and its training can be further optimized by addressing geometric inefficiency. We propose an optimization framework that leverages an asymmetric projection to decompose the backward-pass gradients into parallel spans and orthogonal violations, while keeping the canonical forward-pass $QKV$ structure intact. Through consistent experimental validation across various decomposition and projection setups, we provide strong theoretical evidence: the standard attention gradient is suboptimal. We demonstrated that selectively scaling these components, focusing primarily on $0^{th}$ order bidirectional parallel spans, yields the most effective learning signal. On the limited WikiText-2 dataset, and using a crude configuration, this method achieved a $0.56\%$ reduction in validation loss, confirming the framework's fundamental validity and suggesting significant potential gains on larger datasets and deeper training regimes
format Preprint
id arxiv_https___arxiv_org_abs_2512_13033
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Scaling Bidirectional Spans and Span Violations in Attention Mechanism
Kim, Jongwook
Yun, Sangheon
Yoon, Sukjin
Machine Learning
Artificial Intelligence
Computation and Language
The canonical $O(N^2)$ Transformer remains the empirical performance frontier in sequence modeling, and its training can be further optimized by addressing geometric inefficiency. We propose an optimization framework that leverages an asymmetric projection to decompose the backward-pass gradients into parallel spans and orthogonal violations, while keeping the canonical forward-pass $QKV$ structure intact. Through consistent experimental validation across various decomposition and projection setups, we provide strong theoretical evidence: the standard attention gradient is suboptimal. We demonstrated that selectively scaling these components, focusing primarily on $0^{th}$ order bidirectional parallel spans, yields the most effective learning signal. On the limited WikiText-2 dataset, and using a crude configuration, this method achieved a $0.56\%$ reduction in validation loss, confirming the framework's fundamental validity and suggesting significant potential gains on larger datasets and deeper training regimes
title Scaling Bidirectional Spans and Span Violations in Attention Mechanism
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2512.13033