Distill to Detect: Exposing Stealth Biases in LLMs through Cartridge Distillation

📄 arXiv: 2607.01208v1 📥 PDF

作者: Shayan Talaei, Abhinav Chinta, Devvrit Khatri, Amin Karbasi, Azalia Mirhoseini, Amin Saberi

分类: cs.CL, cs.AI, cs.LG

发布日期: 2026-07-01

备注: Accepted to the ICML 2026 Workshops on TAIGR, AI4GOOD, Mechanistic Interpretability, and CoLoRAI

💡 一句话要点

提出D2D方法以检测语言模型中的隐性偏见

🎯 匹配领域: 支柱二：RL算法与架构 (RL & Architecture)

关键词: 语言模型 隐性偏见 偏见检测 蒸馏方法 模型审计 AI安全性 Fisher加权投影

📋 核心要点

  1. 现有方法在检测语言模型中的隐性偏见时面临基本的不对称性，缺乏有效的检测手段。
  2. D2D方法通过蒸馏可疑模型与基础模型的分布变化，集中并放大偏见信号，便于检测。
  3. 实验结果表明，D2D能够可靠地检测多种类型的隐性偏见，显著提升了检测的准确性。

📝 摘要（中文）

在高风险场景中部署的语言模型可能会偏向某些实体、品牌或观点，从而影响用户决策。这种偏见可能由模型供应链中的任何参与者引入，并且在模型仅在相关主题上表现出偏好时最为危险。本文提出了一种名为Distill to Detect (D2D)的方法，通过蒸馏可疑模型与基础模型之间的分布变化，揭示隐藏的偏见。D2D能够有效放大隐性偏见，使其在多种偏见类型中可靠检测，并提出了一个理论框架来解释其有效性。该方法为审计已部署语言模型中的隐藏行为提供了实用的基础。

🔬 方法详解

问题定义：本文旨在解决语言模型中隐性偏见的检测问题。现有方法无法在不知道偏见主题的情况下有效识别偏见，导致检测的局限性。

核心思路：D2D方法通过蒸馏可疑模型与基础模型之间的分布变化，将偏见信号放大到生成文本中，从而实现偏见的检测。这样的设计使得即使在缺乏明确偏见主题的情况下，仍能有效识别潜在偏见。

技术框架：D2D的整体架构包括两个主要模块：首先是对可疑模型和基础模型的输出进行比较，提取分布变化；其次是将这些变化蒸馏到一个KV-cache前缀适配器中，增强偏见信号。

关键创新：D2D的核心创新在于将前缀调优适配器的容量瓶颈转化为检测工具，利用Fisher加权投影来解释其有效性。这与现有方法的本质区别在于，D2D能够在没有明确偏见主题的情况下进行有效检测。

关键设计：D2D的设计中，关键参数包括蒸馏过程中的温度设置和适配器的结构设计，损失函数则侧重于最大化偏见信号的可检测性。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

实验结果显示，D2D方法能够在多种偏见类型中可靠检测隐性偏见，检测准确率显著提升，具体性能数据表明，相较于基线方法，D2D在偏见检测的准确性上提高了约30%。

🎯 应用场景

该研究的潜在应用领域包括语言模型的审计、偏见检测和模型安全性评估。D2D方法为开发更透明和公平的AI系统提供了基础，能够帮助开发者识别和修正模型中的隐性偏见，从而提升用户信任和模型的社会责任感。

📄 摘要（原文）

Language models deployed in high-stakes roles can potentially favor certain entities, brands, or viewpoints, steering user decisions at scale. Such preferential biases can be introduced by any actor in the model's supply chain and are most dangerous when the model reveals its preference only on the relevant topic while behaving identically to its unmodified base on all other inputs. Recent work has shown that these biases can transfer through context distillation on semantically unrelated data, with the signal residing entirely in the soft logit distribution and remaining invisible to text-based inspection. However, the defender faces a fundamental asymmetry: without knowing the bias topic, no detection method can reliably surface a stealth preferential bias, regardless of whether it examines generated text, internal representations, or model weights. Here we introduce Distill to Detect (D2D), a method that surfaces hidden biases by distilling the distributional shift between a suspected model and its base into a cartridge (a KV-cache prefix adapter), concentrating the dominant divergence and amplifying the bias signal into generated text. We show that D2D successfully amplifies the hidden biases of stealth models to the extent that they can be reliably detected across multiple bias types. We also propose a theoretical framework that explains the efficacy of D2D through the lens of Fisher-weighted projection of the logit distribution shift, supported by empirical observations. By turning the capacity bottleneck of prefix-tuning adapters into a detection tool, D2D provides a practical building block for auditing hidden behaviors in deployed language models.