Abstract
Large Language Models (LLMs) have transformed voice assistants, enabling more natural and versatile conversations. However, the interaction patterns and breakdowns that occur when users converse with these advanced systems remain underexplored. This paper presents a comprehensive analysis of user interactions with LLM-powered voice assistants, identifying common patterns and breakdown points. Through a mixed-methods study combining interaction logs and user interviews, we uncover several key interaction patterns and categorize the types of breakdowns that occur. Our findings reveal how users adapt their interaction strategies based on their mental models of LLM capabilities, and how breakdowns in conversation affect user trust and satisfaction. We provide design implications for improving LLM-powered voice assistants to better support natural conversational interactions.
Dakuo Wang
Associate Professor
My research interests include human-computer interaction, human-AI collaboration, natural language processing, health informatics, data science, and computer-supported cooperative work.