重新思考跨语言对齐：在多语言LLMs中平衡知识迁移与文化抹除
Rethinking Cross-lingual Alignment: Balancing Transfer and Cultural Erasure in Multilingual LLMs
AI总结：
本文提出迁移-定位平面评估框架，发现跨语言对齐虽促进知识迁移但导致文化抹除，并基于不同模型层对两类知识可控性的差异，提出Surgical Steering推理时方法，通过分层激活控制有效平衡了知识迁移与文化定位。
摘要：
跨语言对齐（CLA）旨在对齐多语言表示，使大型语言模型（LLMs）能够无缝地跨语言迁移知识。尽管直观，但我们假设，这种对表示收敛的追求可能会无意中导致“文化抹除”，即无法根据查询语言提供应有所差异的文化情境响应的功能性丧失。在本研究中，我们通过引入一个整体评估框架——迁移-定位平面，系统地分析了这种权衡，该框架同时量化了期望的知识迁移和不应有的文化抹除。使用该框架，我们重新评估了最近的CLA方法，发现它们在所研究的所有六种语言中，始终以直接牺牲文化定位为代价来改善事实迁移。我们对这些模型内部表示的调查揭示了一个关键洞察：普遍的事实迁移和文化特定知识在模型的不同层上具有最佳的可控性。基于这一发现，我们提出了Surgical Steering，一种新颖的推理时方法，将这两个目标解耦。通过对不同层应用有针对性的激活控制，我们的方法在两个竞争维度之间实现了更好的平衡，有效克服了当前对齐技术的局限性。
英文摘要：
Cross-lingual alignment (CLA) aims to align multilingual representations, enabling Large Language Models (LLMs) to seamlessly transfer knowledge across languages. While intuitive, we hypothesize, this pursuit of representational convergence can inadvertently cause "cultural erasure", the functional loss of providing culturally-situated responses that should diverge based on the query language. In this work, we systematically analyze this trade-off by introducing a holistic evaluation framework, the transfer-localization plane, which quantifies both desirable knowledge transfer and undesirable cultural erasure. Using this framework, we re-evaluate recent CLA approaches and find that they consistently improve factual transfer at the direct cost of cultural localization across all six languages studied. Our investigation into the internal representations of these models reveals a key insight: universal factual transfer and culturally-specific knowledge are optimally steerable at different model layers. Based on this finding, we propose Surgical Steering, a novel inference-time method that disentangles these two objectives. By applying targeted activation steering to distinct layers, our approach achieves a better balance between the two competing dimensions, effectively overcoming the limitations of current alignment techniques.
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。
