Prompt Tuning is a lightweight fine-tuning method that demonstrates efficient task adaptation and parameter efficiency for pre-trained language models (PLMs). Prompt Tuning highlights an important contribution to the advancement of NLP technology. The purpose of this paper is to explore the basic principles and methods of Prompt Tuning and to analyze the design features of discrete and continuous Prompts and their application performance in different scenarios. It is found that discrete Prompt performs well in interpretability and simple task adaptation, while continuous Prompt is more advantageous in complex tasks and cross-domain generalization; meanwhile, Prompt Tuning significantly improves the model performance in less sample scenarios, and the parameter efficiency is several times higher than that of traditional fine-tuning. However, the dependence of Prompt Tuning on Prompt design and the limitation of generalization to specific tasks still need to be further optimized. This paper hopes to provide an important reference for future theoretical research and practical applications of Prompt Tuning.
Tongxin Zheng (Wed,) studied this question.