ADF-LoRA: Alternating Low-Rank Aggregation for Decentralized Federated Fine-Tuning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wang, Xiaoyu, Li, Xiaotian, Zhou, Zhixiang, Li, Chen, Liu, Yong
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909919534383104
author Wang, Xiaoyu
Li, Xiaotian
Zhou, Zhixiang
Li, Chen
Liu, Yong
author_facet Wang, Xiaoyu
Li, Xiaotian
Zhou, Zhixiang
Li, Chen
Liu, Yong
contents This paper revisits alternating low-rank updates for federated fine-tuning and examines their behavior in decentralized federated learning (DFL). While alternating the LoRA matrices has been shown to stabilize aggregation in centralized FL, extending this mechanism to decentralized, peer-to-peer communication introduces new challenges due to phase-state mismatch and block-wise divergence across clients. We introduce ADF-LoRA, which synchronizes the update of only one low-rank matrix per round and mixes both matrices to maintain more consistent parameter states under decentralized propagation. This design preserves the cross-term suppression effect of alternating updates while improving stability in serverless topologies. We provide a convergence analysis under standard smoothness assumptions and evaluate ADF-LoRA on multiple GLUE tasks. Experiments show that ADF-LoRA achieves faster and smoother convergence and delivers the highest average accuracy across tasks, outperforming existing LoRA variants in decentralized FL by a consistent margin.
format Preprint
id arxiv_https___arxiv_org_abs_2511_18291
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle ADF-LoRA: Alternating Low-Rank Aggregation for Decentralized Federated Fine-Tuning
Wang, Xiaoyu
Li, Xiaotian
Zhou, Zhixiang
Li, Chen
Liu, Yong
Machine Learning
Distributed, Parallel, and Cluster Computing
I.2.11; I.2.6
This paper revisits alternating low-rank updates for federated fine-tuning and examines their behavior in decentralized federated learning (DFL). While alternating the LoRA matrices has been shown to stabilize aggregation in centralized FL, extending this mechanism to decentralized, peer-to-peer communication introduces new challenges due to phase-state mismatch and block-wise divergence across clients. We introduce ADF-LoRA, which synchronizes the update of only one low-rank matrix per round and mixes both matrices to maintain more consistent parameter states under decentralized propagation. This design preserves the cross-term suppression effect of alternating updates while improving stability in serverless topologies. We provide a convergence analysis under standard smoothness assumptions and evaluate ADF-LoRA on multiple GLUE tasks. Experiments show that ADF-LoRA achieves faster and smoother convergence and delivers the highest average accuracy across tasks, outperforming existing LoRA variants in decentralized FL by a consistent margin.
title ADF-LoRA: Alternating Low-Rank Aggregation for Decentralized Federated Fine-Tuning
topic Machine Learning
Distributed, Parallel, and Cluster Computing
I.2.11; I.2.6
url https://arxiv.org/abs/2511.18291