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 2501.08145cs.CLcs.AI
收藏

大语言模型中的拒绝行为:非线性视角

Refusal Behavior in Large Language Models: A Nonlinear Perspective

FAU Erlangen-Nürnberg(埃尔朗根-纽伦堡大学);University Hospital Erlangen(埃尔朗根大学医院)

Fabian Hildebrandt, Andreas Maier, Patrick Krauss, Achim Schilling

AI总结:

该研究从非线性视角出发,运用PCA、t-SNE、UMAP等降维技术分析三类架构共6个大语言模型的拒绝行为,发现其呈现因架构和层级而异的非线性多维特征,为AI对齐研究与安全部署提供新方向。

摘要:

大语言模型(LLM)的拒绝行为使其能够拒绝对有害、不道德或不当的提示作出回应,确保符合伦理标准。本文研究了来自三个架构家族的六个大语言模型的拒绝行为。我们运用包括PCA、t-SNE和UMAP在内的降维技术,对拒绝行为是线性现象这一假设提出了质疑。研究结果表明,拒绝机制呈现出非线性、多维度的特征,且这些特征因模型架构和层级而异。这些发现凸显了非线性可解释性对于改进对齐研究、为更安全的AI部署策略提供参考的必要性。

英文摘要:

Refusal behavior in large language models (LLMs) enables them to decline responding to harmful, unethical, or inappropriate prompts, ensuring alignment with ethical standards. This paper investigates refusal behavior across six LLMs from three architectural families. We challenge the assumption of refusal as a linear phenomenon by employing dimensionality reduction techniques, including PCA, t-SNE, and UMAP. Our results reveal that refusal mechanisms exhibit nonlinear, multidimensional characteristics that vary by model architecture and layer. These findings highlight the need for nonlinear interpretability to improve alignment research and inform safer AI deployment strategies.

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。

↑