Persona-Pruner: Sculpting Lightweight Models for Role-Playing

The paper introduces Persona-Pruner, a framework designed to optimize language models for role-playing applications by isolating persona-specific sub-networks, thereby reducing computational costs. Experiments demonstrate that Persona-Pruner minimizes performance degradation compared to traditional pruning methods, achieving a reduction in performance drop by up to 93.8% on the RoleBench benchmark while retaining the general capabilities of the original model. This advancement is significant for practitioners as it allows for more efficient deployment of LLMs in environments with multiple non-player characters (NPCs), enhancing scalability and performance.