Extensible Prompts for Language Models on Zero-shot Language Style Customization

We propose eXtensible Prompt (X-Prompt) for prompting a large language model (LLM) beyond natural language (NL). X-Prompt instructs an LLM with not only NL but also an extensible vocabulary of imaginary words. Imaginary words can help represent what NL words hardly describe, allowing a prompt to be more descriptive; also, they are designed to be out-of-distribution (OOD) robust so that they can be used like NL words in various prompts, distinguishing X-Prompt from soft prompt that is for fitting in-distribution data. To this end, we propose context-augmented learning (CAL) to learn imaginary words for general usability, enabling them to work properly in OOD (unseen) prompts. We conduct experiments that use X-Prompt for zero-shot language style customization as a case study. The promising results of X-Prompt demonstrate its potential of approaching advanced interaction between humans and LLMs to bridge their communication gap.