A Survey of Large Language Model Architectures and their Impact on Machine Translation

A Survey of Large Language Model Architectures and their Impact on Machine TranslationAuthor: Zarmeen Ahmed Overview This survey reviews the role of large language models (LLMs) in advancing machine translation (MT). It synthesizes recent developments in model architectures (transformers, encoder–decoder systems, foundation models), training strategies (zero-shot, few-shot, fine-tuning, domain adaptation), and data techniques (synthetic data generation, multilingual pretraining). The paper compares LLM-based approaches with traditional MT methods, highlights performance trends across high and low-resource language pairs, and discusses evaluation challenges (automatic metrics, human evaluation, robustness). It concludes with open challenges and recommended research directions for creating more accurate, robust, and inclusive MT systems. Contributions Systematic categorization of LLM architectures relevant to MT. Comparative discussion of prompting, fine-tuning, and adaptation strategies. Analysis of LLM performance tradeoffs for low-resource and multilingual scenarios. Practical recommendations for evaluation protocols, dataset creation, and future research directions. License & reusePublished under Creative Commons Attribution 4.0 (CC BY 4.0). Please cite the work for reuse; contact the author for permissions on large excerpts or figure reuse beyond the license. Publication status & IEEE notePreprint / conference-format survey. If accepted for publication in an IEEE conference, this record will be updated with the official IEEE citation and DOI and the accepted manuscript will be uploaded in accordance with IEEE post-publication policies. How to citeZarmeen Ahmed (2025). A Survey of Large Language Model Architectures and their Impact on Machine Translation. Zenodo. DOI: 10.5281/zenodo.17076209