[论文] Persona-Pruner: Sculpting Lightweight Models for Role-Playing

论文概要

研究领域: ML 作者: Jinsu Kim, Jihoon Tack, Noah Lee 发布时间: 2026-06-12 arXiv: 2606.14695

中文摘要

语言模型（LMs）作为角色扮演聊天机器人显示出显著潜力，在提供角色或用户人设规范时能够交付一致、风格化的交互。然而，将这些能力应用于现实世界应用（例如，大量NPC同时交互的生态系统）暴露了一个关键低效性，由于过高的计算成本。在本文中，我们质疑将完整、通用模型专用于单一人设的必要性，假设特定角色身份仅依赖于模型总容量的一小部分。我们观察到，朴素地剪枝LMs往往会严重降低特定人设的角色扮演性能；它无法区分冗余知识和 essential 角色特征。我们提出了Persona-Pruner，一个通过从单一描述中隔离人设特定子网络来塑造轻量级角色扮演模型的框架。我们的实验一致表明，Persona-Pruner比现有最先进的LLM剪枝技术更有效地保持角色扮演性能，在RoleBench的LLM-as-a-judge评分上比最强基线减少性能下降高达93.8%，同时仍保持通用LLM能力。代码在 https://github.com/jsu-kim/Persona-Pruner 可用。

原文摘要

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 Pers...

--- *自动采集于 2026-06-16*

#论文 #arXiv #ML #小凯