PEFT-Arena: Understanding Parameter-Efficient Finetuning from a Stability-Plasticity Perspective

arXiv:2605.28819 · 2026-05-28 공개 · arXiv · PDF

peft stability-plasticity parameter-efficient-finetuning representation-distortion singular-value-analysis path-wise-rewinding general-capability-retention

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.

한국어 요약

📋 한 줄 요약

**[PEFT 평가 / Stability-Plasticity]** PEFT-Arena가 downstream 성능과 일반 capability 보존을 함께 평가 — orthogonal finetuning이 동일 budget에서 가장 favorable Pareto frontier, weight·activation 기하 분석·path-wise rewinding 후처리 제시.

🎯 핵심 기여도

💡 핵심 아이디어

PEFT 평가는 stability-plasticity dilemma로 재정의해야 하며, weight space의 spectral interaction과 activation space의 non-isometric representation distortion이 forgetting의 기하학적 원인이며 — 동일 budget에서는 orthogonal finetuning이 가장 균형 잡힌 Pareto frontier를 차지하고, SFT overshoot은 path-wise rewinding으로 post-hoc 회복 가능하다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: PEFT 평가 framework의 패러다임 전환, 기하학적 원인 해명, orthogonal finetuning의 우수성 정량 입증, 실용적 post-hoc 보정(rewinding) 제시. **한계**: 평가 대상 PEFT 방법 범위, weight·activation 기하 분석의 모델·task 의존성, path-wise rewinding이 case study 수준.

🚀 실용적 활용