Get Started
Home
Topics
Search
Library
Research questionHow can Arabic LLMs generate accurate target dialects from MSA prompts without fine-tuning?Arabic LLMs receive far less dialectal data than Modern Standard Arabic, so they often default to MSA instead of producing an accurate dialect. It remains unclear whether dialectal information is localized in a few internal features or distributed across the model, and how that affects controllability.
AI
Mechanistic Interpretability
Natural Language Processing
Research Paper
Latest papersRecent research connected to this question, newest first.Can Dialects Be Steered Like Languages? Sparse Neurons and Distributed Directions in Arabic LLMsThe evidence concerns Arabic LLMs and inference-time interpretability probes without fine-tuning. Sparse neuron populations occupy under 1% of MLP dimensions and cover only 5–21% of the residual dialect direction; manipulating them reinforces dialect when prompts are already dialectal but does not induce dialect from MSA prompts, while distributed activation-direction steering succeeds in both settings.research paper · Sep 6, 2026
Related questions
Can large language models reliably perform Arabic morphosyntactic tagging and dependency parsing despite morphological and orthographic ambiguity?How can prompt representations distinguish specificity to expose fine-grained LLM weaknesses?How can we build reproducible Armenian LLMs without sacrificing knowledge for fluency?How can grammar-constrained decoding preserve syntactic validity without distorting an LLM’s output probabilities?