当模型拒绝时：政治可操控性与特征丰富度作为意识形态深度的度量
When Models Refuse: Political Steerability and Feature Richness as Measures of Ideological Depth
AI总结：
本文提出意识形态深度概念，通过可操控性和稀疏自编码器测量的特征丰富度，研究大语言模型拒绝遵循良性指令是否源于能力缺陷而非安全规则。
摘要：
大型语言模型有时会拒绝遵循良性指令，例如拒绝论证某种政治立场或采用指定角色，这种拒绝通常被视为安全护栏在起作用。我们探究这些拒绝是否可能表明**能力缺陷**：模型缺乏从指令角度进行推理所需的内部表示。为此，我们引入**意识形态深度**，该属性包含两个组成部分：(i) 模型遵循政治指令而不*失败*的能力（可操控性），以及 (ii) 其内部政治表示的**特征丰富度**，通过稀疏自编码器测量。使用两个广泛使用的开源权重LLM作为候选，我们比较基于提示和激活操控的干预措施，并用公开可用的SAE探测政治特征。我们发现巨大且系统性的差异：在两个意识形态方向上更具可操控性的模型激活了约**7.3倍**更多的独特政治特征，而另一个模型则以增加拒绝作为回应。从前者模型中因果性地消融一小部分目标政治特征，重现了相同的特征贫乏行为并导致拒绝增加。综合这些结果表明，对良性提示的拒绝可能源于**能力缺陷**而非固定的安全规则，并且意识形态深度是LLM的一个可测量属性，有助于预测模型何时会拒绝。
英文摘要：
Large language models (LLMs) sometimes refuse to follow benign instructions, such as declining to argue a political position or adopt a stated persona, and such refusals are commonly read as safety guardrails at work. We ask whether they can instead signal a **capability deficit**: a shortage of the internal representations a model needs to reason from the instructed perspective. To investigate, we introduce **ideological depth**, a property with two components: (i) a model's ability to follow political instructions without *failure* (steerability), and (ii) the **feature richness** of its internal political representations, measured with sparse autoencoders (SAEs). Using two widely used openweight LLMs as candidates, we compare interventions based on prompts and activation-steering, and probe political features with publicly available SAEs. We find large, systematic differences: a model that is more steerable in both ideological directions activates **~7.3x** more distinct political features, while the other model instead responds with increased refusals. Causally ablating a small, targeted set of political features from the former model reproduces the same feature-poor behavior and drives up refusals. Together, these results indicate that refusals on benign prompts can arise from **capability deficits** rather than fixed safety rules, and that ideological depth is a measurable property of LLMs that helps predict when a model will refuse.
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。
