SPARKLING: Balancing Signal Preservation and Symmetry Breaking for Width-Progressive Learning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Yu, Qifan, Ma, Xinyu, Zhuo, Zhijian, Wang, Minrui, Liu, Deyi, Zhan, Shiyi, Ma, Yiyuan, Xiang, Liang, Bin, Xingyan, He, Di
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915768957927424
author Yu, Qifan
Ma, Xinyu
Zhuo, Zhijian
Wang, Minrui
Liu, Deyi
Zhan, Shiyi
Ma, Yiyuan
Xiang, Liang
Bin, Xingyan
He, Di
author_facet Yu, Qifan
Ma, Xinyu
Zhuo, Zhijian
Wang, Minrui
Liu, Deyi
Zhan, Shiyi
Ma, Yiyuan
Xiang, Liang
Bin, Xingyan
He, Di
contents Progressive Learning (PL) reduces pre-training computational overhead by gradually increasing model scale. While prior work has extensively explored depth expansion, width expansion remains significantly understudied, with the few existing methods limited to the early stages of training. However, expanding width during the mid-stage is essential for maximizing computational savings, yet it remains a formidable challenge due to severe training instabilities. Empirically, we show that naive initialization at this stage disrupts activation statistics, triggering loss spikes, while copy-based initialization introduces gradient symmetry that hinders feature diversity. To address these issues, we propose SPARKLING (balancing {S}ignal {P}reservation {A}nd symmet{R}y brea{K}ing for width-progressive {L}earn{ING}), a novel framework for mid-stage width expansion. Our method achieves signal preservation via RMS-scale consistency, stabilizing activation statistics during expansion. Symmetry breaking is ensured through asymmetric optimizer state resetting and learning rate re-warmup. Extensive experiments on Mixture-of-Experts (MoE) models demonstrate that, across multiple width axes and optimizer families, SPARKLING consistently outperforms training from scratch and reduces training cost by up to 35% under $2\times$ width expansion.
format Preprint
id arxiv_https___arxiv_org_abs_2602_02472
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle SPARKLING: Balancing Signal Preservation and Symmetry Breaking for Width-Progressive Learning
Yu, Qifan
Ma, Xinyu
Zhuo, Zhijian
Wang, Minrui
Liu, Deyi
Zhan, Shiyi
Ma, Yiyuan
Xiang, Liang
Bin, Xingyan
He, Di
Machine Learning
Computation and Language
Progressive Learning (PL) reduces pre-training computational overhead by gradually increasing model scale. While prior work has extensively explored depth expansion, width expansion remains significantly understudied, with the few existing methods limited to the early stages of training. However, expanding width during the mid-stage is essential for maximizing computational savings, yet it remains a formidable challenge due to severe training instabilities. Empirically, we show that naive initialization at this stage disrupts activation statistics, triggering loss spikes, while copy-based initialization introduces gradient symmetry that hinders feature diversity. To address these issues, we propose SPARKLING (balancing {S}ignal {P}reservation {A}nd symmet{R}y brea{K}ing for width-progressive {L}earn{ING}), a novel framework for mid-stage width expansion. Our method achieves signal preservation via RMS-scale consistency, stabilizing activation statistics during expansion. Symmetry breaking is ensured through asymmetric optimizer state resetting and learning rate re-warmup. Extensive experiments on Mixture-of-Experts (MoE) models demonstrate that, across multiple width axes and optimizer families, SPARKLING consistently outperforms training from scratch and reduces training cost by up to 35% under $2\times$ width expansion.
title SPARKLING: Balancing Signal Preservation and Symmetry Breaking for Width-Progressive Learning
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2602.02472