Prompt-Calibrated Tuning: Improving Black-Box Optimization for Few-Shot Scenarios

When there is limited downstream data available, using prompts can still yield good results. With prompts, large language models(LLMs) have already achieved success in numerous NLP tasks. Therefore, it is necessary to find the optimal prompts for each task. Due to the limitation of only having access to LLMs' API without the ability to perform backpropagation, we optimize the prompts using derivative-free optimization. In this paper, we propose Prompt-Calibrated Tuning(PCT), which modifies the outputs of large language models before prompt tuning and keeps their parameters frozen. Besides, we adopt the Roberta-large model as the backbone model, which might split a word into several subwords during tokenization, resulting in insufficient masking. Therefore, we utilize whole-word mask(wwm) in this paper. In addition, we have also expanded the label words. The experimental results show that Prompt-Calibrated Tuning is more effective than existing black-box optimization.