User Modeling Challenges in Interactive AI Assistant Systems

TL;DR

This paper tackles the challenge of inferring users' mental states to personalize guidance in interactive AI assistants. It extends the WTaG dataset with six user-profile categories and analyzes correlations and cross-task consistency across 55 recordings from 17 users and 3 recipes. Using prompted ChatGPT to predict these categories from dialog history, the study finds high accuracy for Detail-oriented, Eager to ask questions, and Talkative, but substantial false positives for Frustration and poor detection for Experienced, indicating gaps. These results reveal both the potential and limitations of current LLM-based user modeling and suggest directions such as larger datasets, refined prompting, online detection, and multimodal cues to improve personalization.

Abstract

Interactive Artificial Intelligent(AI) assistant systems are designed to offer timely guidance to help human users to complete a variety tasks. One of the remaining challenges is to understand user's mental states during the task for more personalized guidance. In this work, we analyze users' mental states during task executions and investigate the capabilities and challenges for large language models to interpret user profiles for more personalized user guidance.

Figures (3)

• Figure 1: Feature Correlations
• Figure 2: User Mental State Consistency Across Recipes
• Figure 3: ChatGPT User Modeling Prediction Performance