Least but not Last: Fine-tuning Intermediate Principal Components for Better Performance-Forgetting Trade-Offs

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Quercia, Alessio, Bangun, Arya, Assent, Ira, Scharr, Hanno
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866918321235951616
author Quercia, Alessio
Bangun, Arya
Assent, Ira
Scharr, Hanno
author_facet Quercia, Alessio
Bangun, Arya
Assent, Ira
Scharr, Hanno
contents Low-Rank Adaptation (LoRA) methods have emerged as crucial techniques for adapting large pre-trained models to downstream tasks under computational and memory constraints. However, they face a fundamental challenge in balancing task-specific performance gains against catastrophic forgetting of pre-trained knowledge, where existing methods provide inconsistent recommendations. This paper presents a comprehensive analysis of the performance-forgetting trade-offs inherent in low-rank adaptation using principal components as initialization. Our investigation reveals that fine-tuning intermediate components leads to better balance and show more robustness to high learning rates than first (PiSSA) and last (MiLoRA) components in existing work. Building on these findings, we provide a practical approach for initialization of LoRA that offers superior trade-offs. We demonstrate in a thorough empirical study on a variety of computer vision and NLP tasks that our approach improves accuracy and reduces forgetting, also in continual learning scenarios.
format Preprint
id arxiv_https___arxiv_org_abs_2602_03493
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Least but not Last: Fine-tuning Intermediate Principal Components for Better Performance-Forgetting Trade-Offs
Quercia, Alessio
Bangun, Arya
Assent, Ira
Scharr, Hanno
Machine Learning
Low-Rank Adaptation (LoRA) methods have emerged as crucial techniques for adapting large pre-trained models to downstream tasks under computational and memory constraints. However, they face a fundamental challenge in balancing task-specific performance gains against catastrophic forgetting of pre-trained knowledge, where existing methods provide inconsistent recommendations. This paper presents a comprehensive analysis of the performance-forgetting trade-offs inherent in low-rank adaptation using principal components as initialization. Our investigation reveals that fine-tuning intermediate components leads to better balance and show more robustness to high learning rates than first (PiSSA) and last (MiLoRA) components in existing work. Building on these findings, we provide a practical approach for initialization of LoRA that offers superior trade-offs. We demonstrate in a thorough empirical study on a variety of computer vision and NLP tasks that our approach improves accuracy and reduces forgetting, also in continual learning scenarios.
title Least but not Last: Fine-tuning Intermediate Principal Components for Better Performance-Forgetting Trade-Offs
topic Machine Learning
url https://arxiv.org/abs/2602.03493