GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Choudhary, Anand, Sulaıman, Yasser, Mauch, Lukas, Hacene, Ghouthi Boukli, Cardinaux, Fabien, Bosselut, Antoine
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914108243181568
author Choudhary, Anand
Sulaıman, Yasser
Mauch, Lukas
Hacene, Ghouthi Boukli
Cardinaux, Fabien
Bosselut, Antoine
author_facet Choudhary, Anand
Sulaıman, Yasser
Mauch, Lukas
Hacene, Ghouthi Boukli
Cardinaux, Fabien
Bosselut, Antoine
contents Sparse fine-tuning techniques adapt LLMs to downstream tasks by only tuning a sparse subset of model parameters. However, the effectiveness of sparse adaptation depends on optimally selecting the model parameters to be fine-tuned. In this work, we introduce a novel sparse fine-tuning technique named GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters, which fine-tunes only those model parameters which have the largest gradient magnitudes on downstream tasks and the smallest pre-trained magnitudes, intuitively prioritizing parameters that are highly task-relevant, but minimally disruptive to pre-trained knowledge. Our experimentation with LLaMA3 8B and Gemma 2B as base models shows that GaLLoP consistently improves or matches the in-distribution as well as out-of-distribution performance obtained via the usage of other leading parameter-efficient fine-tuning techniques, including LoRA, DoRA, and SAFT. Our analysis demonstrates that GaLLoP mitigates catastrophic forgetting and memorization of task data, as important pre-trained parameters remain unchanged, and stabilizes performance relative to other fine-tuning techniques, robustly generalizing across most random seeds.
format Preprint
id arxiv_https___arxiv_org_abs_2510_19778
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters
Choudhary, Anand
Sulaıman, Yasser
Mauch, Lukas
Hacene, Ghouthi Boukli
Cardinaux, Fabien
Bosselut, Antoine
Machine Learning
Computation and Language
Sparse fine-tuning techniques adapt LLMs to downstream tasks by only tuning a sparse subset of model parameters. However, the effectiveness of sparse adaptation depends on optimally selecting the model parameters to be fine-tuned. In this work, we introduce a novel sparse fine-tuning technique named GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters, which fine-tunes only those model parameters which have the largest gradient magnitudes on downstream tasks and the smallest pre-trained magnitudes, intuitively prioritizing parameters that are highly task-relevant, but minimally disruptive to pre-trained knowledge. Our experimentation with LLaMA3 8B and Gemma 2B as base models shows that GaLLoP consistently improves or matches the in-distribution as well as out-of-distribution performance obtained via the usage of other leading parameter-efficient fine-tuning techniques, including LoRA, DoRA, and SAFT. Our analysis demonstrates that GaLLoP mitigates catastrophic forgetting and memorization of task data, as important pre-trained parameters remain unchanged, and stabilizes performance relative to other fine-tuning techniques, robustly generalizing across most random seeds.
title GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2510.19778