Adapting virtual agent interaction style with reinforcement learning to enhance affective engagement

The ability of artificial agents to dynamically adapt their communication style is a key factor in sustaining engagement during human-agent interaction. This study introduces a reinforcement learning-based framework for real-time modulation of interaction style, aiming to maximize the affective valence of the user's emotional response. The approach is domain-independent and designed for integration into scenarios where personalized and engaging dialogue is critical, such as in Behavior Change Interventions.

To validate the system, we conducted a between-subjects user study involving $N=20$ participants, who completed a structured task, i.e. the URICA questionnaire, delivered either by an adaptive speech-based agent or a static screen-based interface. In the adaptive condition, the virtual agent employed Thompson Sampling to select between two communication styles (enthusiastic and neutral) based on real-time facial emotion recognition. The goal of the system was to reinforce the style that increased or maintained valence across successive interaction turns.

The reinforcement learning system successfully adapted its behavior based on individual users' emotional feedback. Notably, a significant positive correlation was observed between users' Psychoticism scores and the reinforcement of the neutral style (Spearman's $\rho =0.70$ , p-value = 0.04), indicating sensitivity to personality traits. Although no significant differences emerged in user-reported experience between conditions, this highlights that the adaptive speech-based agent preserved usability while successfully personalizing interaction based on affective cues.

These findings highlight the potential of adaptive agents to personalize interaction strategies in emotionally relevant contexts, even when the subjective user experience appears similar to that of static systems. The ability to align communicative behavior with user personality profiles supports the feasibility of deploying such models in long-term interventions, where maintaining user motivation and engagement is essential.