Large Language Models for Natural Language Processing: Architectures, Training Paradigms, and Real-World Applications – A Systematic Review

Abstract

In this manuscript, Large Language Models (LLMs) are praised as a transformative paradigm in the scope of man-made insight, which invigorates tremendous improvement in the capacities of Natural Language Processing (NLP) mechanisms. The triumph of transformer architectures and large scale pre- training has lead to latest LLMs to consistently provide strong performance for a wide range of tasks - from generating text to question answering, translation and reasoning - thus demonstrating their true utility. The ability of these models for assimilation of intricate datasets through linguistic pattern recognition has led to a great deal of progress in both academic research and industrial practice. This investigation presents a holistic and systematic look at the recent scholarship on large language models, and it has a concentrated focus on the topics of architectural evolution, training paradigms, and real-world applications. A systematically structured literature search was carried out from the most prominent academic databases: IEEEXplore database, ACM Digital Library, Scopus, Science Direct and arXiv from 2023 to 2026. The study follows a methodology that is modelled on Preferred Reporting Items for Systematic Reviews and Meta,-Analyses (PRISMA) guidelines. Items for Systematic Reviews and Meta-Analyses (PRISMA), and included the use of well-defined inclusion and exclusion criteria to identify relevant peer reviewed articles and high impact preprints. The review explains important new developments in LLM architectures, which include improvements in transformer architectures, mixture-of-experts architectures, and new alternatives which are focused on efficient sequence modelling. It further challenges the paradigms of training, such as training from a large scale (pre-training), instruction tuning, parameter efficient fine tuning (e.g. LoRA), reinforcement learning from human feedback (RLHF). In addition to this, the paper highlights the growing use of LLMs in other areas such as healthcare, education, software, and scientific research. In spite of their successes, there are significant challenges remaining - hallucination, bias, computational cost and evaluation limitations. The open research problems and possible futures for efficient, reliable, and trustworthy large language models are identified, ending the review.