Prompt Optimization for Personalized Response Generation

Abstract

Large language models (LLMs) have demonstrated remarkable performance capabilities. However, their limited open-source accessibility restricts general users from adjusting the internal parameters of the models. Consequently, generating personalized responses with LLMs requires the careful design of prompts. This paper proposes a novel automated prompt optimization method that generates and stores knowledge for prompt optimization and reuses it in future response generation. Our approach consists of two key components. First, it selects examples which are similar to the current task to include in the prompt, and it determines whether selected examlpes should be included in the prompt or not. Second, it generates insights for determining which examples should be included in a prompt. Experimental results demonstrate that prompts generated using the proposed method achieve significantly higher response accuracy compared to prompts without examples.