In the context of parameter-efficient fine-tuning, adaptor layers inserted between existing model layers can be viewed as a type of soft prompt. These layers encode prompting and task-specific information, interacting with the original, frozen Large Language Model to facilitate its adaptation to new tasks.