PAIR: Perceptual Affective Inference and Regulation in a Real-Time Multimodal Conversational Agent

In plain English

A real-time conversational agent estimated a user's emotional state from the conversation, then delivered emotion-regulation guidance through coordinated speech, colour and avatar cues, carrying context across sessions with a rolling memory. Across 19 participants and 1,093 sessions over 14 days, its valence and dominance estimates were moderately accurate (mean absolute error 1.20 and 1.30 on a 9-point scale) while arousal agreed only weakly. The useful finding for practitioners is the gap between the two: users' sense of being understood predicted mood improvement, while numerical prediction error barely did — so getting the affect numbers right and making someone feel understood are not the same engineering problem.