Abstract
The advent of Large Language Models (LLMs) has significantly reshaped thelandscape of machine translation (MT), particularly for low-resource languagesand domains that lack sufficient parallel corpora, linguistic tools, andcomputational infrastructure. This survey presents a comprehensive overview ofrecent progress in leveraging LLMs for MT. We analyze techniques such asfew-shot prompting, cross-lingual transfer, and parameter-efficient fine-tuningthat enable effective adaptation to under-resourced settings. The paper alsoexplores synthetic data generation strategies using LLMs, includingback-translation and lexical augmentation. Additionally, we compare LLM-basedtranslation with traditional encoder-decoder models across diverse languagepairs, highlighting the strengths and limitations of each. We discusspersistent challenges such as hallucinations, evaluation inconsistencies, andinherited biases while also evaluating emerging LLM-driven metrics fortranslation quality. This survey offers practical insights and outlines futuredirections for building robust, inclusive, and scalable MT systems in the eraof large-scale generative models.