Does Modern Standard Arabic (MSA) Dominate Arabic Dialects in LLMs? A Representation-Level Analysis

Key points

• Large language models (LLMs) often default to Modern Standard Arabic (MSA) when generating Arabic, even when prompted with dialectal Arabic.1
• Instead, dialect representations form a dense and highly overlapping space: normalized mutual information drops sharply compared to patterns reported for more distinct languages.1
• These findings challenge a common interpretation of MSA-biased generation: output preference does not necessarily reveal internal representational dominance.1
• Analyses of multilingual and dialectal LLMs should therefore distinguish generation bias from the geometry of internal representations.1

Sources (1)

• [1]Does Modern Standard Arabic (MSA) Dominate Arabic Dialects in LLMs? A Representation-Level Analysis
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 08:46 AM

  Across layers and model families, we find no evidence that MSA acts as a dominant internal representation for Arabic dialects.

  Large language models (LLMs) often default to Modern Standard Arabic (MSA) when generating Arabic, even when prompted with dialectal Arabic.

Extractive summary: sentences quoted from the sources.