Cross-Lingual Transfer and Parameter-Efficient Adaptation in the Turkic Language Family: A Theoretical Framework for Low-Resource Language Models

This work presents a structured approach to developing and evaluating large language model (LLM) systems for low-resource languages, with a specific focus on Azerbaijani. While recent advances in LLMs have significantly improved multilingual capabilities, performance in underrepresented languages remains inconsistent and often lacks robust evaluation frameworks. This paper introduces a methodology for improving language performance through targeted model adaptation, evaluation design, and system-level optimization. The work combines elements of model fine-tuning, retrieval-augmented generation (RAG), and domain-specific benchmarking to create a practical and scalable framework for real-world deployment. Key contributions: A structured evaluation framework for Azerbaijani language performance across domains Practical approach to adapting foundation models (e.g., Qwen family) for low-resource settings Integration of RAG pipelines for improving factual accuracy and domain grounding Empirical results demonstrating measurable improvements in downstream tasks This research is particularly relevant for government, enterprise, and public-sector applications where language-specific accuracy, auditability, and reliability are critical. arXiv preprint:https://arxiv.org/abs/2604.06202 Keywords: Azerbaijani language model, Low-resource NLP, Large Language Models (LLM), Multilingual AI, Qwen models, RAG (Retrieval-Augmented Generation), AI for government, Speech and language processing, Domain adaptation, Evaluation benchmarks, ALLMAZ - Advanced Large Language Model for AZerbaijan.