Persona-Pruner: Sculpting Lightweight Models for Role-Playing

Forum topic · 小凯 · 2026-06-16

Summary

Persona-Pruner is a framework for creating lightweight role-playing language models by isolating persona-specific subnetworks from a single character description. The authors observe that naive pruning of LMs severely degrades persona-specific role-playing performance because it cannot distinguish redundant knowledge from essential character traits. Persona-Pruner addresses this by sculpting compact models tailored to individual personas, reducing the need to dedicate a full generalist model to one character. Experiments show it preserves role-playing performance better than state-of-the-art LLM pruning techniques, cutting performance degradation by up to 93.8% on RoleBench LLM-as-a-judge scores versus the strongest baseline, while maintaining general LLM capabilities. This makes it well suited for real-world applications such as ecosystems with many simultaneously interacting NPCs, where computational cost is critical. Paper by Jinsu Kim, Jihoon Tack, and Noah Lee (arXiv:2606.14695), with code available on GitHub.

Paper Overview

Field: ML Authors: Jinsu Kim, Jihoon Tack, Noah Lee Published: 2026-06-12 arXiv: 2606.14695

Abstract

Language Models (LMs) have shown remarkable potential as role-playing chatbots, delivering consistent, stylized interactions when given a character or user persona specification. 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 lightweight role-playing models by isolating persona-specific subnetworks from a single description.

Our experiments consistently demonstrate that Persona-Pruner preserves role-playing performance more effectively than existing state-of-the-art LLM pruning techniques, reducing performance degradation by up to 93.8% in LLM-as-a-judge scores on RoleBench compared to the strongest baseline, while still maintaining general LLM capabilities.

Code: https://github.com/jsu-kim/Persona-Pruner

---

*Auto-collected on 2026-06-16*

Persona-Pruner: Sculpting Lightweight Models for Role-Playing

Summary

Paper Overview

Abstract

Tags