Enhancing Persona Following at Decoding Time via Dynamic Importance Estimation for Role-Playing Agents

Researchers propose a new Persona Dynamic Decoding (PDD) framework that enables AI role-playing agents to dynamically adapt their personas based on context during inference time. The method uses psychological theories to estimate persona importance and adjust behavior without requiring expensive fine-tuning or static prompts.

• →Current role-playing AI agents struggle to adapt personas dynamically to different scenarios, limiting their realism in social simulations.
• →The new PDD framework includes a Persona Importance Estimation module that quantifies contextual relevance of persona attributes without supervision.
• →The system uses weighted multi-objective rewards to modulate generation probabilities during inference rather than requiring costly model retraining.
• →The approach is grounded in psychological theory, specifically Cognitive-Affective Personality Systems, explaining why persona influence varies by context.
• →Experimental results demonstrate improved utterance consistency and behavioral fidelity in role-playing language agents.