Persona-Pruner: Sculpting Lightweight Models for Role-Playing

2026-06-12

AI Summary

This work addresses the prohibitive computational cost of deploying large language models (LLMs) in multi-NPC role-playing scenarios. To this end, we propose Persona-Pruner, a novel framework that introduces, for the first time, identity-aware structured pruning tailored to specific character personas. Leveraging character descriptions, our method automatically identifies and extracts dedicated subnetworks, substantially compressing the model while preserving both essential persona-specific traits and general language capabilities. Unlike conventional generic pruning strategies, Persona-Pruner achieves markedly superior performance on the RoleBench benchmark, incurring only a 6.2% performance drop relative to the strongest baseline—equivalent to a 93.8% reduction in performance loss compared to existing LLM pruning approaches.

Abstract

Language Models (LMs) have shown remarkable potential as role-playing chatbots, delivering consistent, stylized interactions when given a specification of a character or user persona. However, applying these capabilities to real-world applications (e.g., ecosystems with numerous NPCs interacting simultaneously) exposes a critical inefficiency due to the excessive computational cost. In this paper, we question the necessity of dedicating a full, generalist model to a single persona, hypothesizing that a specific character identity relies on only a fraction of the model's total capacity. We observe that naively pruning LMs often severely degrades the role-playing performance for a specific persona; it does not distinguish between redundant knowledge and essential character traits. We propose Persona-Pruner, a framework that sculpts a lightweight role-playing model by isolating persona-specific sub-networks from a single description. Our experiments consistently show that Persona-Pruner preserves role-playing performance substantially more effectively than existing state-of-the-art LLM pruning techniques, reducing the performance drop from the dense model by up to 93.8% over the strongest baseline on RoleBench in LLM-as-a-judge score, while still maintaining general LLM capabilities. Code is available at https://github.com/jsu-kim/Persona-Pruner.

Problem

• role-playing
• large language models
• model pruning
• computational efficiency
• persona-specific

Innovation

• Persona-Pruner
• role-playing LLMs
• model pruning
• persona-specific subnetworks
• lightweight language models