Search papers, labs, and topics across Lattice.
This paper introduces PEFT-Arena, a benchmark to evaluate parameter-efficient finetuning (PEFT) methods based on the stability-plasticity dilemma, measuring both downstream performance and general capability retention. They find that orthogonal finetuning achieves the best Pareto frontier under comparable parameter budgets and that forgetting is linked to non-isometric representation distortion in activation space. Further, they demonstrate that standard SFT checkpoints often overshoot the optimal target-retention balance, and can be improved via path-wise rewinding.
PEFT methods aren't just about downstream accuracy; they have distinct "stability-plasticity profiles" that reveal how well they retain general capabilities, and most overshoot the optimal balance anyway.
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.
