Abstract: To address the problems of high computational cost, heavy storage requirements, and low deployment efficiency of full fine-tuning in adapting PLMs ( Pretrained Language Models) to downstream tasks, the development of PEFT(Parameter-Efficient Fine-Tuning) techniques is studied, and a classification framework covering adapter-based methods, low-rank adaptation methods, prompt-based methods, selective fine-tuning methods, and dynamic tuning methods is constructed. These methods are comparatively analyzed in terms of their basic principles, representative approaches, applicable scenarios, and application characteristics. The analysis shows that different PEFT methods exhibit distinct advantages in parameter efficiency, task generalization ability, and deployment flexibility, and can provide references for the lightweight deployment, multi-task transfer, and personalized adaptation of large models. 

Key words: parameter-efficient fine-tuning (PEFT), adapter-based methods, low-rank adaptation, prompt- based methods, selective fine-tuning methods, dynamic tuning methods