Towards Personalized AI: Engineering Model Responses for Customized User Interactions in Generative AI Systems

Yoonsu Kim, Junhyeon Park

Abstract: Recent advancements in generative AI models, particularly the emergence of ChatGPT, have significantly increased interest in generative AI. However, the reliance on text prompts as the primary medium of interaction poses various challenges, especially for users with limited AI knowledge. Prompt engineering has emerged as a technique to address this issue, but it is still time-consuming and may not always yield satisfactory results. In this research, we aim to test the feasibility of an AI system that provides customized answers tailored to individual users, focusing on engineering model responses rather than solely relying on prompt engineering. We developed a system that enables users to evaluate their preferences and provide feedback on several model responses for the same prompt. And through this system, we collected user preference data and applied instruction tuning to guide model responses in a manner preferred by a certain user based on the certain user's preference data. Our study serves as a proof of concept to explore the potential of more personalized and user-centric AI systems in the future.