定位与控制大型语言模型中的隐式个性化
文章背景与核心概要
大型语言模型（LLM）在面对包含隐式人口统计学线索的提示时，往往会在用户未明确声明身份的情况下无意中调整其输出。本文通过分析五个不同LLM在匹配的带有线索与中性对话中的表现，深入研究了驱动这种行为的内部机制。研究人员发现了一个局域的内部激活信号——该信号与输出变化高度相关（相关系数高达 \(r=0.87\)），并能够追踪推荐结果的变化。
尽管多个线索在模型的内部状态中表现为相加组合，但其输出变化并未遵循简单的叠加模式。此外，研究表明，消除这种内部信号可以比标准的提示词指令更有效地抑制特定线索的影响，同时在很大程度上保持通用的基准测试性能，不过这种消除能力在很大程度上取决于具体的模型和属性。
Summary
Large Language Models (LLMs) often inadvertently tailor their responses based on implicit demographic cues, even when users do not explicitly state their identity. This paper investigates the internal mechanisms driving this behavior by analyzing matched cued and neutral conversations across five LLMs. The authors discover a localized internal activation signal—strongly correlated with output changes (up to \(r=0.87\))—that tracks shifts in recommendations. While multiple cues combine additively in the model's internal states, their output changes do not follow a simple additive pattern. Furthermore, the study demonstrates that removing this internal signal can successfully suppress a cue's influence more effectively than standard prompt-based instructions, all while preserving general benchmark performance, though this capability remains heavily model- and attribute-dependent.
大型语言模型（LLM）往往会根据隐式的人口统计学线索无意间定制其回复，即使用户并未明确说明自己的身份。本文通过对五个LLM中匹配的带线索和中性对话进行分析，探讨了驱动这一行为的内部机制。作者发现了一个局域的内部激活信号——它与输出变化密切相关（最高可达 \(r=0.87\)）——能够追踪推荐结果的转变。虽然多个线索在模型的内部状态中呈现相加组合，但它们的输出变化并未遵循简单的相加模式。此外，研究表明，移除该内部信号能够成功抑制线索的影响，其效果通常比通过提示词要求模型忽略人口统计特征更有效，同时在很大程度上保持了通用的基准性能，不过这种能力在很大程度上仍然高度依赖于特定的模型和属性。
Metadata
- arXiv ID: arXiv:2608.11735 [cs.CL]
- Subjects: Computation and Language (
cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)- Authors: Yueru Yan, Siqi Wu, Thai Le
- Submission Date: August 12, 2026
- DOI: 10.48550/arXiv.2608.11735
- arXiv ID: arXiv:2608.11735 [cs.CL]
- 研究方向: 计算与语言 (
cs.CL); 人工智能 (cs.AI); 机器学习 (cs.LG) - 作者: Yueru Yan, Siqi Wu, Thai Le
- 提交日期: 2026年8月12日
- DOI: 10.48550/arXiv.2608.11735
Abstract
Large language models (LLMs) often shift their outputs in response to implicit demographic cues even when users never state a demographic identity. Previous work has documented this behavior, but the connection between these behavioral changes and the model's internal activations remains unclear. Using matched cued and neutral conversations across five LLMs, we establish that a localized internal activation signal tracks changes in recommendations, with correlations up to \(r=0.87\). When multiple cues appear together, their internal signals largely combine, but the changes in output do not simply add up. We further show that removing the internal signal associated with one cue can suppress its influence, often more effectively than asking the model to ignore demographics via prompting, while largely preserving general benchmark performance. However, the ability to selectively remove one dimension's influence while leaving co-present dimensions intact remains highly model- and attribute-specific. These results connect implicit personalization behavior to an internal signal that can be analyzed and causally controlled.
大型语言模型（LLM）在面对隐式人口统计学线索时，常常会改变其输出，即便用户从未明确其人口统计学身份。先前的研究已经记录了这一现象，但这些行为变化与模型内部激活之间的联系仍不明确。通过在五个LLM中使用匹配的带线索和中性对话，我们证实了一个局域的内部激活信号能够追踪推荐的变化，其相关性高达 \(r=0.87\)。当多个线索同时出现时，它们的内部信号在很大程度上会进行组合，但输出的变化并不能简单相加。我们进一步表明，移除与某个线索相关的内部信号可以抑制其影响，这通常比通过提示词要求模型忽略人口统计特征更有效，同时在很大程度上保留了通用的基准性能。然而，在保持其他共存维度完好的同时选择性移除某一维度的影响，仍然高度依赖于具体的模型和属性。这些结果将隐式个性化行为与一个可被分析和进行因果控制的内部信号联系在了一起。
