Exploring The Principles and Prospects for Efficient Fine-Tuning of Transformer-Based Pre-Trained Large Language Models

In recent years, large language models (LLMs) have made breakthroughs in natural language processing and multimodal tasks. However, the growing model size and the high cost of full parameter fine-tuning pose challenges to their efficient adaptation. This paper focus on Transformer-based Parameter Efficient Fine-Tuning (PEFT) techniques for large models, and analyze three types of methods, namely, additive-based, specification-based and reparameterization-based, from the perspectives of performance, engineering complexity and applicability. This paper concludes that the PEFT technique exhibits comparable or even better performance than full parameter fine-tuning in a variety of tasks, but still faces stability and adaptability challenges in complex scenarios. Future research can further advance the field by improving flexibility, optimizing strategies and focusing on privacy and security. The purpose of this paper is to provide a basic reference for researchers to understand the PEFT algorithm and its system implementation, to further promote the implementation of PEFT technology in the research industry.