语言模型中的文化绑定头
Cultural Binding Heads in Language Models
AI总结：
通过机制可解释性和析因设计，识别出8个语言模型中2-3个中间层注意力头对文化绑定有因果贡献，且绑定主要在预训练阶段形成，知识探测表明模型知道的知识远多于其行为表现。
摘要：
大型语言模型通常默认对不同文化群体一视同仁，即使上下文需要区分：这缺乏差异意识。利用机制可解释性和Wang等人(2025)的N4文化挪用基准上的析因设计，我们在八个模型（四种架构，基础版和指令版）中识别出每个模型有2-3个中间层注意力头对文化绑定有因果贡献。文化绑定是将文化项目与适当身份关联的过程。敲除这些头上的身份到项目边会使绑定强度降低9-23%。识别出的头从指令模型转移到基础模型，表明文化绑定是在预训练阶段创建的。α缩放显示分级剂量反应，生成时适度放大引导（α=2-3）可将文化区分准确性提高1-3个百分点，同时基本保持中性推理不变。知识探测任务表明，模型知道的知识比其行为表现多3-5倍，表明瓶颈在于路由而非知识。
英文摘要：
LLMs often default to equal treatment across cultural groups, even though context warrants differentiation: this is a lack of difference awareness. Using mechanistic interpretability and a factorial design on the N4 cultural appropriation benchmark from Wang et al. (2025), we identify 2-3 mid-layer attention heads per model that contribute causally to cultural binding across eight models (base and instruct versions of four architectures). Cultural binding is the process of associating a cultural item with its related identity. Knockout of the identity-to-item edges on these heads lowers the binding strength by 9-23%. The identified heads transfer from instruct to base models, suggesting that cultural binding is created during pre-training. An $α$-scaling shows a graded dose-response. Moderate amplification steering at generation ($α= 2-3$) increases cultural differentiation accuracy by 1-3 pp while leaving reasoning on culturally neutral questions mostly intact. A knowledge probing task shows that models know 3-6 times more than they act upon, indicating that the bottleneck lies in routing and not knowledge.
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。
