What is prompt tuning and where is it applied?

Prompt tuning is a method of adapting large language models (LLMs) without full retraining of the model. Instead of changing all the weights of the model, prompt tuning trains a small number of additional parameters that act as 'prompts', influencing the model's behavior.

Prompt tuning is used when there is a need to quickly and efficiently adapt an LLM for a specific task, such as text classification, response generation, or translation, with limited resources. This is especially useful when there is no possibility or desire to fine-tune the entire model due to its size or cost.

Example: instead of full fine-tuning of a GPT model, one can train only special vector representations (soft prompts), which are fed into the model and guide its generation in the desired direction.