Researchers unveil CPM-MultiAgent for emotional AI characters that evolve in dialogue

A new research paper introduces CPM-MultiAgent, a framework that advances persona-based dialogue agents by enabling dynamic emotional evolution in AI characters. Unlike existing systems that treat emotions as static traits or surface-level cues, this approach models a character's emotional state as a latent variable continuously reshaped by triggers in multi-turn conversations. Drawing from the Component Process Model (CPM) from psychology, the system breaks down emotional responses into a multi-step process: extracting affective triggers from dialogue, collaboratively appraising them through multiple agents, and updating the emotion state accordingly.

Extensive evaluations including baseline comparisons, ablation studies, and human assessments show that CPM-MultiAgent produces more emotionally consistent and believable role simulations. The framework is particularly relevant for applications in healthcare counseling, education, therapy, customer service, and interactive storytelling where realistic emotional progression is critical. By grounding AI emotional dynamics in established psychological theory, this work bridges a key gap between persona-based dialogue systems and affective computing.