From Triggers to Emotions: A CPM-Grounded Appraisal Multi-Agent for Dynamic Emotional Evolution in Persona-Based Dialogue

Persona-based dialogue systems often treat emotions as fixed attributes or stylistic cues, failing to account for how a character's internal emotional state shifts in response to external dialogue triggers. This paper addresses the challenge of modeling dynamic, trigger-driven emotional evolution in role-simulation scenarios, such as healthcare training, counseling, and customer service.

The authors propose CPM-MultiAgent, a framework inspired by the Component Process Model (CPM) of psychology. Instead of a monolithic model, the framework uses a multi-agent architecture to decompose the emotional process into distinct, logical stages:

1. Trigger Analysis: Extracts affectively salient stimuli from dialogue turns and grounds them in the character's persona and context.
2. CPM Appraisal: Four specialized agents evaluate triggers across four dimensions: relevance, implications, coping potential, and normative significance. These agents engage in peer review to ensure the appraisal chain is coherent.
3. Integration: An agent synthesizes these appraisals to update the character's latent emotional state.
4. Consistency Critique: A final agent audits the entire process for faithfulness to the CPM theory and contextual consistency before the emotion state is used to generate a response.

Experiments across healthcare, education, and customer service scenarios demonstrate that CPM-MultiAgent outperforms standard prompting and existing agent-based baselines. By explicitly modeling the appraisal process, the framework achieves more emotionally consistent and psychologically grounded role simulation. The multi-agent design effectively separates complex reasoning tasks, reducing the tendency of LLMs to produce formulaic or overly positive responses, and allows for more nuanced, multi-dimensional emotional transitions over multi-turn interactions.