跳转至

通过非局域性测量 SAE 特征的语义抽象度

文章背景与核心概要

大语言模型（LLM）的内部机制研究正越来越多地依赖稀疏自编码器（SAE）来解构诸如复杂推理和越狱等行为背后的任务相关特征。然而，在实际研究中，如何区分表层词汇特征与真正的高阶语义概念始终是一大挑战，因为现有的自动解释（autointerp）描述或因果引导效用都无法完全准确地界定特征的抽象级别。

为了解决这一痛点，作者引入了一种名为特征非局域性（Feature Nonlocality, FNL）的新指标。该指标通过计算归一化后每个位置对 SAE 特征激发的相对影响的熵来定义。研究表明，FNL 提供了一种独立于大模型本身、且无需人工标注的语义抽象度测量方式，能够成功将依赖上下文的推理特征与由特定 Token 驱动的特征区分开来。这项工作为评估大模型机制解释以及下游干预中的特征筛选提供了坚实的理论支持和有效工具。

测量语义抽象度 via 非局域性 (Measuring Semantic Abstractness of SAE Features via Nonlocality)

摘要 (Summary)

Sparse autoencoders (SAEs) help uncover mechanistic explanations for Large Language Model (LLM) behaviors—such as reasoning and jailbreaking—by isolating task-relevant features. However, distinguishing between surface-level lexical features and genuinely high-level semantic concepts remains challenging, as neither autointerp descriptions nor causal steering utilities fully capture a feature's abstraction level.

稀疏自编码器（SAE）通过隔离任务相关的特征，有助于揭示大语言模型（LLM）行为（如推理和越狱等）的机制性解释。然而，区分表层词汇特征与真正的高阶语义概念仍然具有挑战性，因为无论是基于自动解释（autointerp）的语义描述还是因果引导效用，都无法完全捕捉特征的抽象级别。

To address this, the authors introduce Feature Nonlocality (FNL), a metric defined as the entropy of the normalized per-position influence on an SAE feature's activation. FNL provides an LLM-independent, label-free way to measure semantic abstractness, successfully separating context-dependent reasoning features from token-driven ones.

为了解决这一问题，作者引入了特征非局域性（Feature Nonlocality, FNL），该指标定义为归一化后每个位置对 SAE 特征激发的相对影响的熵。FNL 提供了一种独立于 LLM 且无需标签的方法来测量语义抽象度，能够成功将上下文相关的推理特征与 Token 驱动的特征区分开来。

文章详情 (Article Details)

  • arXiv ID: 2608.10537
  • Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
  • Authors: Chuqiao Lin, Shivaji Sondhi, Xiao-Liang Qi
  • Submitted: August 11, 2026
  • Citations / Formats: View PDF | HTML Version
  • arXiv ID: 2608.10537
  • 研究领域: 人工智能 (cs.AI)；机器学习 (cs.LG)
  • 作者: Chuqiao Lin, Shivaji Sondhi, Xiao-Liang Qi
  • 提交时间: 2026年8月11日
  • 引用 / 格式: 查看 PDF | HTML 版本

摘要正文 (Abstract)

Sparse autoencoders (SAEs) have helped uncover mechanistic explanations for LLM behaviours such as reasoning, jailbreaking etc., via understanding the corresponding task-relevant and causally effective features. To evaluate such mechanistic explanations, downstream studies must distinguish surface lexical features from genuinely high-level ones. However, neither an autointerp-based semantic description nor causal steering utility fully resolves the abstraction level of a feature.

稀疏自编码器（SAE）通过理解相应的任务相关和因果有效特征，帮助揭示了 LLM 行为（如推理、越狱等）的机制解释。为了评估此类机制解释，下游研究必须将表层词汇特征与真正的高阶特征区分开来。然而，基于自动解释的语义描述和因果引导效用都无法完全确定特征的抽象级别。

To this end, we introduce Feature Nonlocality (FNL), defined as the entropy of the normalized per-position influence on an SAE feature's activation. We report that FNL correlates with existing LLM-based proxy metrics of feature semantic abstractness, and successfully distinguishes context-dependent reasoning features from token-driven ones, correctly assigning the higher FNL to the contextual feature in \(73\)--\(84\%\) of randomly drawn pairs that consist of one contextual and one token-level feature.

为此，我们引入了特征非局域性（Feature Nonlocality, FNL），其定义为归一化后每个位置对 SAE 特征激活的相对影响的熵。我们报告称，FNL 与现有的基于 LLM 的特征语义抽象度代理指标相关，并且能够成功区分上下文相关的推理特征与 Token 驱动的特征；在由一个上下文特征和一个 Token 级特征组成的随机抽取特征对中，FNL 在 \(73\%\) 至 \(84\%\) 的情况下能正确将更高的 FNL 赋予上下文特征。

关键下游应用 (Key Downstream Applications)

  1. Jailbreak Mitigation Auditing: The authors audited SAE features used for jailbreak mitigation, discovering that the most effective features are typically positional features with low FNL, rather than features that genuinely recognize harmful intents.
  2. Model Steering and Performance Gains: Steering high-FNL features in DeepSeek-R1-Distill-Llama-8B improves MATH-500 accuracy by \(4.6\) points over the unsteered model, outperforming the steering of low-FNL features (though these gains remain model-specific).
  1. 越狱缓解审计： 作者审计了用于越狱缓解的 SAE 特征，发现最有效的特征通常是具有低 FNL 的位置特征，而不是真正识别有害意图的特征。
  2. 模型引导与性能提升： 在 DeepSeek-R1-Distill-Llama-8B 中引导高 FNL 特征，可使 MATH-500 的准确率比未引导模型提升 \(4.6\) 个百分点，其效果优于引导低 FNL 特征（尽管这些增益仍具模型特异性）。

Conclusion: FNL provides a robust, correlational witness to the abstraction level of an SAE feature, offering valuable utility for evaluating mechanistic explanations and selecting features for downstream interventions.

结论： FNL 为 SAE 特征的抽象级别提供了一个强大的、基于相关的佐证指标，为评估机制解释以及选择用于下游干预的特征提供了宝贵的实用价值。

访问与资源 (Access and Resources)

(Note: Original license icon reference retained for compliance: license icon)

(注：为合规起见，保留原始许可证图标引用：license icon)