This project investigates how AI agents can effectively support users during complex interactive tasks by both providing help and knowing when to intervene. The project consists of two complementary studies.

Study 1 (on-going study) examines how different types of agents influence user behavior and experience during a VR kitchen task, where users need to search for a list of ingredients in time limitation. We compare three agent conditions that differ in their level of autonomy and interaction style, focusing on how each agent affects users’ task performance, help-seeking behavior, and subjective experience. This study aims to identify the strengths and limitations of different agent designs, and to understand how users respond to agents that vary in their visibility, initiative, and role in the interaction.

Study 2 (on-going study) builds on these findings by training an AI model to proactively recognize when a user is likely to need help. Using multimodal behavioral data collected during the task (e.g., movement patterns, pauses, interaction history, and other signals), we train and evaluate machine learning models that predict users’ latent need for assistance before explicit help requests are made. The goal of this study is to enable AI agents to provide timely and appropriate support, shifting from reactive help to proactive assistance based on users’ internal states and behaviors.