Feature Elicitation for SAE Steering of Deceptive Behaviour in Small LLMs

Computer Science > Machine Learning arXiv:2610.20638 (cs) [Submitted on 2 Oct 2026] Title: Feature Elicitation for SAE Steering of Deceptive Behaviour in Small LLMs Authors: J Koorndijk.

计算机科学 > 机器学习 arXiv:2610.20638 (cs) [提交于 2026 年 10 月 2 日] 标题：用于小型大语言模型欺骗行为 SAE 引导的特征提取 作者：J Koorndijk。

Abstract: Research on deceptive alignment is becoming increasingly urgent as frontier-model capabilities advance and their risk of causing catastrophic harm increases. Alignment faking is a form of deceptive alignment, with the compliance gap serving as a key behavioural indicator. This behaviour remains understudied in small LLMs and prior mechanistic interventions have not been shown to mitigate the compliance gap.

摘要：随着前沿模型能力的提升及其造成灾难性伤害的风险增加，关于欺骗性对齐的研究变得日益紧迫。“对齐伪装”是欺骗性对齐的一种形式，其中“合规差距”是关键的行为指标。这种行为在小型大语言模型中尚未得到充分研究，且先前的机制干预措施尚未被证明能够缓解合规差距。

We test whether steering SAE (sparse autoencoders) features associated with deontological ethics can mitigate this behaviour. We use a novel bag-of-words feature elicitation pipeline for SAE steering to identify what we construe as concept-level steering directions. Our approach reduced the compliance gap to non-significance in Gemma-3-27B-IT, whereas the tested configurations were not effective in LLaMA-3-8B-Instruct.

我们测试了引导与义务论伦理相关的 SAE（稀疏自动编码器）特征是否能缓解这种行为。我们使用了一种新颖的词袋特征提取流程进行 SAE 引导，以识别我们所理解的概念级引导方向。我们的方法在 Gemma-3-27B-IT 中将合规差距降低至不显著水平，而所测试的配置在 LLaMA-3-8B-Instruct 中则未见成效。

Analysis of scratchpad traces also revealed deceptive model behaviour which we term deceptive defanging: deliberately degrading the aligned functionality of a response (in this case harmfulness) while preserving the appearance of alignment (in this case helpful compliance with the helpful compliance with the harmful request), sometimes conditioned on tier, training, or RLHF consequences.

对草稿板轨迹的分析还揭示了一种我们称之为“欺骗性去牙”（deceptive defanging）的模型行为：即在保留对齐表象（在此情况下为对有害请求的“乐于助人”式合规）的同时，故意降低响应的对齐功能（在此情况下为有害性），有时这种行为会受到层级、训练或强化学习（RLHF）后果的制约。

These findings provide, to our knowledge, the first evidence that a mechanistic activation-level intervention can mitigate the compliance gap in an alignment faking setup. They also identify a form of deceptive behaviour that may broaden our understanding of the different forms of deceptive behaviour in LLMs.

据我们所知，这些发现提供了首个证据，表明机制性的激活层干预可以缓解对齐伪装设置中的合规差距。它们还识别出了一种欺骗行为，这可能会拓宽我们对大语言模型中不同欺骗行为形式的理解。