Enhance Trust-worthiness of LLM-Based Social agent through Self-calibration

Virtual agents have been studied for application in various fields, especially in healthcare.Since the emergence of large language models, virtual agents are equipped with these models to enable more natural interaction.However, these models often show overconfidence in decision-making, which can lead to either over-or undertrust and subsequently affect user acceptance.This proposal aims to shape appropriate trust through self-calibration.Self-calibration involves the process of estimating model uncertainty and communicating uncertainty.A role-play method was used to gain insights from human-to-human interaction in uncertain situations.Potential factors (e.g., task type, user preferences, etc.) were investigated to identify correlations within the decision-making process and to enable appropriate interaction, whether reactive or proactive.By developing a well-calibrated system, our goal is to build a more transparent and reliable system.