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

2026-05-27

AI Summary

This work addresses a critical gap in the evaluation of parameter-efficient fine-tuning (PEFT) methods, which typically prioritize downstream task performance while neglecting the preservation of pre-trained models’ general capabilities and lacking systematic analysis of the trade-off between stability (resistance to forgetting) and plasticity (task adaptability). To this end, we introduce PEFT-Arena, the first benchmark framework that jointly assesses task performance and capability retention. Leveraging weight-space spectral analysis, activation-space representation fidelity, and non-isometric distortion detection, our framework enables a deeper mechanistic understanding of PEFT behaviors. Experiments reveal that orthogonal fine-tuning occupies the Pareto-optimal frontier in the stability–plasticity trade-off under identical parameter budgets, whereas standard supervised fine-tuning often yields suboptimal balances due to over-optimization. Furthermore, our proposed path-backtracking strategy significantly enhances the final checkpoint’s retention–performance trade-off.

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.

Problem

• Parameter-Efficient Finetuning
• Stability-Plasticity Dilemma
• Catastrophic Forgetting
• Model Adaptation
• Pretrained Capability Retention

Innovation

• parameter-efficient finetuning
• stability-plasticity dilemma
• PEFT-Arena
• spectral analysis
• representation distortion