Persona-Pruner: Sculpting Lightweight Role-Playing Models via Persona-Specific Subnetworks

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—Jinsu Kim, Jihoon Tack, and Noah Lee—observe that serving a full generalist model for each persona is computationally inefficient in real-world deployments such as ecosystems with many simultaneously interacting NPCs, while naive pruning severely degrades role-playing quality because it fails to distinguish redundant knowledge from essential character traits. Persona-Pruner addresses this by identifying and preserving only the parameters relevant to a given character identity. Experiments show it retains role-playing performance more effectively than state-of-the-art LLM pruning techniques, reducing performance degradation by up to 93.8% on RoleBench LLM-as-a-judge scores compared to the strongest baseline, while maintaining general LLM capabilities. The paper is available on arXiv (2606.14695) and code is released at github.com/jsu-kim/Persona-Pruner.

Paper Overview

Field: Machine Learning
Authors: Jinsu Kim, Jihoon Tack, Noah Lee
Published: 2026-06-12
arXiv: 2606.14695
Code: https://github.com/jsu-kim/Persona-Pruner

Abstract (Translation)

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 excessive computational cost. In this paper, the authors 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.

A key observation is that naively pruning LMs often severely degrades role-playing performance for a specific persona: pruning does not distinguish between redundant knowledge and essential character traits.

The proposed Persona-Pruner framework sculpts lightweight role-playing models by isolating persona-specific subnetworks from a single character description. Experiments consistently show 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% on RoleBench LLM-as-a-judge scores versus the strongest baseline, while still maintaining general LLM capabilities.

Key Takeaways

A full generalist LLM is unnecessary to serve a single persona; character identity depends on a small subset of model capacity.
Naive pruning conflates redundant knowledge with essential persona traits, hurting role-play quality.
Persona-Pruner extracts persona-specific subnetworks directly from a character description.
Up to 93.8% lower performance degradation than the best pruning baseline on RoleBench (LLM-as-a-judge).
General LLM abilities are retained after pruning.