PEFT-Arena: Understanding Parameter-Efficient Finetuning from a Stability-Plasticity Perspective
2026-05-27 • Machine Learning
Machine LearningComputation and Language
AI summaryⓘ
The authors discuss a popular way to update large language models called parameter-efficient finetuning (PEFT). They say that people usually check how well updated models do on new tasks but forget to see if models keep their original skills. To fix this, the authors made a test called PEFT-Arena that looks at both learning new things and remembering old ones. They found different PEFT methods balance this trade-off differently and explain why by studying how updates change the model inside. They also suggest a way to improve these updates after training using something called path-wise rewinding.
parameter-efficient finetuninglarge language modelsstability-plasticity dilemmageneral capability retentionorthogonal finetuningspectral analysisweight spaceactivation spacerepresentation distortionpath-wise rewinding
Authors
Yangyi Huang, Ruotian Peng, Zeju Qiu, Jiale Kang, Yandong Wen, Bernhard Schölkopf, Weiyang Liu
Abstract
Parameter-efficient finetuning (PEFT) has become the standard approach for adapting large language models, yet evaluations largely emphasize downstream accuracy while overlooking the retention of pretrained capabilities. We argue that PEFT should be assessed through the stability-plasticity dilemma: the trade-off between target-task adaptation and resistance to forgetting. We introduce PEFT-Arena, a benchmark that jointly measures downstream performance and general capability retention. Across methods, we find distinct stability-plasticity profiles; under comparable parameter budgets, orthogonal finetuning achieves the most favorable Pareto frontier. To explain these differences, we analyze PEFT updates from two geometric perspectives. In weight space, spectral analysis reveals how parameterizations interact with the pretrained singular-value structure. In activation space, retention metrics show whether finetuning preserves or distorts general-capability representations, with forgetting linked to non-isometric representation distortion. Finally, an analysis shows that final SFT checkpoints often overshoot a better target-retention operating point. Inspired by this, we present case studies of a post-hoc improvement with path-wise rewinding.