LLM 条件控制中效果与流畅度的权衡研究
原标题：On the Effectiveness-Fluency Trade-Off in LLM Conditioning: A Systematic Study
AI 摘要
Apple 机器学习研究团队发表了一项系统性研究，考察大语言模型条件控制（conditioning）中效果与流畅度之间的权衡。研究发现，高效的激活引导（activation steering）方法往往以牺牲生成流畅度为代价，且在指令微调模型上的效果远不如基座模型。简单提示和监督微调适合概念注入但不擅长概念移除，同时低成本文本指标与昂贵的 LLM-as-judge 评分高度相关。
正文节选
On the Effectiveness-Fluency Trade-Off in LLM Conditioning: A Systematic Study AuthorsIuri Macocco†, Pau Rodríguez Lopez, Arno Blaas, Luca Zappella, Marco Baroni†*, Xavier Suau Cuadros* Controlling the output of Large Language Models (LLMs) is a central challenge for their reliable deployment, yet a clear understanding of the involved trade-offs remains elusive. Current approaches to conditioning are often evaluated with a narrow focus on their effectiveness at injecting or removing a target con