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