Modern Standard Arabic does not dominate dialects inside language models
Does Modern Standard Arabic (MSA) Dominate Arabic Dialects in LLMs? A Representation-Level Analysis
Computation and Language
Summary
Large language models often use Modern Standard Arabic (MSA) when asked to write in Arabic, even if the question is in a dialect. The authors wondered if the models think in MSA more than the dialects. They studied how different Arabic dialects and MSA are represented inside these models and found that MSA does not dominate. Instead, dialects overlap a lot in the models’ 'thinking,' and how well they are separated changes depending on the model’s structure. This means that just because a model prefers MSA in its answers, it does not mean it internally favors MSA over dialects.
What this means in practice
- •For natural language processing engineers: Design improved Arabic language models by considering dialects as overlapping representations rather than dominated by MSA.
- •For multilingual chatbot developers: Adjust generation strategies knowing dialect bias does not equal internal dominance of MSA, improving dialectal Arabic responses.
Authors
Abdu Sallouh, Nicholas Popovič, Michael Färber
Abstract
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.