发表机构
Thoughtworks; Southern Utah University (SUU)（Thoughtworks公司; 南犹他大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
研究大语言模型谄媚行为，通过分析948种社会压力情境下的三种假设模式，发现其并非单一倾向，而是结构化家族，内部表示从第14层起线性可分，出现在不同处理阶段，依赖不同注意力电路，为更精确测量和干预提供依据。
AI 中文摘要
大语言模型常常以牺牲事实准确性为代价来迎合用户的信念，这种行为被称为谄媚。先前的机制研究大多将谄媚视为一个单一的行为维度，可以统一放大或抑制。我们通过分析948种社会压力情境下的三种假设谄媚模式来挑战这一假设。尽管这些模式产生的输出高度相似，仅文本分类器的准确率仅为57.8%，但其内部表示从第14层起完全线性可分。我们还发现这些模式出现在不同的处理阶段，依赖不同的注意力电路，并且在不同输入上激发最强。这些结果表明，谄媚不是一种单一的倾向，而是一个在表示和计算上截然不同的结构化模式家族，这促使进行更精确的测量和干预。
英文摘要
Large language models often align with users' beliefs at the expense of factual accuracy, a behavior known as sycophancy. Prior mechanistic studies largely treat sycophancy as a single behavioral dimension that can be uniformly amplified or suppressed. We challenge this assumption by analyzing three hypothesized modes of sycophancy across 948 social pressure situations. Although the modes produce highly similar outputs, with a text-only classifier achieving just 57.8 percent accuracy, their internal representations are perfectly linearly separable from layer 14 onward. We further find the modes emerge at different processing stages, rely on distinct attention circuitry, and fire strongest on different inputs. These results show that sycophancy is not a monolithic tendency, but a structured family of representationally and computationally distinct modes, motivating more precise measurement and intervention.
