LoRA vs Full Fine-tuning: An Illusion of Equivalence

Reece Shuttleworth, Jacob Andreas, Antonio Torralba, Pratyusha Sharma

arXiv:2410.21228 · 2026-07-27 공개 · arXiv · PDF

large-language-models continual-learning lora singular-value-decomposition forgetting full-finetuning spectral-properties intruder-dimensions

Abstract

Fine-tuning is a crucial paradigm for adapting pre-trained large language models to downstream tasks. Recently, methods like Low-Rank Adaptation (LoRA) have been shown to effectively fine-tune LLMs with an extreme reduction in trainable parameters. But, \emph{are their learned solutions really equivalent?} We study how LoRA and full-finetuning change pre-trained models by analyzing the model's weight matrices through the lens of their spectral properties. We find that LoRA and full fine-tuning yield weight matrices whose singular value decompositions exhibit very different structure: weight matrices trained with LoRA have new, high-ranking singular vectors, which we call \emph{intruder dimensions}, while those trained with full fine-tuning do not. Further, we extend the finding that LoRA forgets less than full fine-tuning and find its forgetting is vastly localized to the intruder dimension -- by causally intervening on the intruder dimensions by changing their associated singular values post-fine-tuning, we show that they cause forgetting. Moreover, scaling them down significantly improves modeling of the pre-training distribution with a minimal drop in downstream task performance. Given this, we should expect accumulating intruder dimensions to be harmful and lead to more forgetting. This will be amplified during continual learning because of sequentially fine-tuning, and we show that LoRA models do accumulate intruder dimensions here tend to perform worse in this setting, emphasizing the practicality of our findings.

한국어 요약

한 줄 요약

LoRA와 full fine-tuning은 유사한 성능을 보이지만, 학습된 솔루션의 스펙트럼 구조와 일반화 능력에 차이가 있음을 밝힘.

핵심 기여도

핵심 아이디어

LoRA는 low-rank 매트릭스의 곱으로 가중치 업데이트를 표현함. 이는 full fine-tuning보다 훈련 파라미터 수를 극적으로 줄이지만, 학습된 솔루션의 구조적 차이가 존재함. 본 연구는 SVD를 통해 LoRA와 full fine-tuning이 pre-trained 모델의 가중치 행렬에 미치는 영향을 분석함. LoRA 모델은 pre-trained 모델과 직교한 고유한 singular vector, 즉 intruder dimensions를 생성하며, 이는 pre-training 분포를 잊는 주요 원인임. 이는 LoRA가 full fine-tuning과 동일한 솔루션을 학습하지 않음을 시사함. 또한, LoRA의 rank가 높을수록 full fine-tuning과 유사한 성능을 보이지만, rank-stabilization 없이 높은 rank를 사용하면 일반화 능력이 저하됨.

기술적 접근법

주요 결과

의의 및 한계

실용적 활용