发表机构
Technical University of Dresden; Mohamed bin Zayed University of Artificial Intelligence（德累斯顿工业大学; 穆罕默德·本·扎耶德人工智能大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本研究通过调整语言主导框架分析26种阿拉伯语变体，发现LLMs中MSA并非阿拉伯方言的内部主导表示，输出偏好与内部表示主导性无必然关联，需区分二者开展相关分析。
AI 中文摘要
大语言模型（LLMs）在生成阿拉伯语文本时，即便提示使用阿拉伯方言，也常默认采用现代标准阿拉伯语（MSA）。一个合理的解释是，模型的内部表示由MSA主导。我们通过调整Shani和Basirat（2025）的语言主导框架（该链接为https URL），对26种阿拉伯语变体开展测试。在各层及不同模型家族中，我们未发现MSA作为阿拉伯方言内部主导表示的证据；相反，方言表示形成了密集且高度重叠的空间，与差异更显著语言的相关模式相比，归一化互信息急剧下降。此外，最强的可分性效应并不局限于中间层，而是会根据架构向后续层转移。这些发现对MSA生成偏见的常见解释提出挑战：输出偏好未必反映内部表示的主导性，因此多语言及方言LLMs的分析应区分生成偏见与内部表示的几何结构。
英文摘要
Large language models (LLMs) often default to Modern Standard Arabic (MSA) when generating Arabic, even when prompted with dialectal Arabic. A natural explanation is that their internal representations are dominated by MSA. We test this hypothesis by adapting the language-dominance framework of Shani and Basirat (2025) (https://doi.org/10.18653/v1/2025.blackboxnlp-1.7) to 26 Arabic varieties. Across layers and model families, we find no evidence that MSA acts as a dominant internal representation for Arabic dialects. Instead, dialect representations form a dense and highly overlapping space: normalized mutual information drops sharply compared to patterns reported for more distinct languages. Moreover, the strongest separability effects are not confined to intermediate layers, but can shift toward later layers depending on the architecture. These findings challenge a common interpretation of MSA-biased generation: output preference does not necessarily reveal internal representational dominance. Analyses of multilingual and dialectal LLMs should therefore distinguish generation bias from the geometry of internal representations.
CommentsEMNLP 2026
