arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.02343cs.CLcs.AIcs.CVcs.IRcs.LG
收藏

为何转向有效:语言模型参数动态的统一视角

Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics

Zhejiang University(浙江大学);Alibaba Group(阿里巴巴集团)

Ziwen Xu, Chenyan Wu, Hengyu Sun, Haiwen Hong, Mengru Wang, Yunzhi Yao, Longtao Huang, Hui Xue, Shumin Deng, Zhixuan Chu, Huajun Chen, Ningyu Zhang

AI总结:

本文提出统一视角分析语言模型控制方法,通过偏好-效用分析揭示控制效果的权衡关系,并引入SPLIT方法提升偏好与效用。

摘要:

针对大型语言模型(LLM)控制方法的研究常孤立进行,掩盖其联系并使比较困难。本文提出统一视角,将这些干预视为由控制信号诱导的动态权重更新,置于单一概念框架内。基于此视角,我们提出统一的偏好-效用分析,将控制效果分为偏好(倾向目标概念)和效用(一致且任务有效生成),并在共享的对数几率尺度上使用极性配对对比示例进行测量。在各种方法中,我们观察到偏好与效用之间的一致权衡:更强的控制增加偏好但预测性降低效用。我们进一步通过激活流形视角解释此行为,其中控制沿目标概念方向移动表示以增强偏好,而效用下降主要当干预将表示推离模型的有效生成流形时。最后,我们引入一种新的转向方法SPLIT,通过此分析提升偏好并更好地保持效用。代码可在https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md获取。

英文摘要:

Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑