Learning process when a new task come

[CuongNN218]

Hi authors, Thanks for your great work. I have a question related to learning process for when a different task come which indicated in the Section 3.2.1 in your paper. In your paper, you said that "we include an additional set of prompt tokens for the new task and fine-tune their embeddings while keeping the old ones frozen". In my opinion, I have 2 different understanding about this sentence. The 1st is that when training with a new task, you concatenate the new set of prompts for the current tasks (learnable) with the old prompts from previous tasks (frozen) and the input for training. The new prompt at time t is initialized by the prompt from time t-1. My second assumption is that for a new task at time t, you only init new prompt by the prompt from time t-1 and concatenate it with the input. Here, all old prompts are still frozen and not concatenated with the input and the new prompt. Which one is correct? Please help me to clarify my confusion. Best.