Large Language Models: The Core Of Generative AI — A Review Of Architecture, Training, Capabilities And Limitations

Authors: Nimse Vidya Kisan, Prof. J. M. Patade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Large Language Models (LLMs) have become the core technology of Generative Artificial Intelligence, powering chat assistants, writing and coding tools, search and many other applications. This paper reviews how LLMs work and why they matter. It explains the transformer architecture with self-attention, tokenization into subword units and the decoder-only design used by generative models, then traces the development from generative pre-training through models that perform tasks from a few examples. Scaling laws and compute-optimal training explain why larger models trained on more data perform better, and emergent abilities and chain-of-thought prompting show how capabilities appear with scale. Alignment through instruction tuning and reinforcement learning from human feedback, and grounding through retrieval, are reviewed. Open model families are discussed. A five-stage LLM lifecycle architecture, comprising data, pre-training, alignment, augmentation and application, is proposed. Research gaps, objectives, a comparison of model families and limitations including hallucination, bias, cost, environmental impact, privacy and misuse are analysed. The review concludes that LLMs are a foundational general-purpose technology whose benefits depend on responsible development, grounding and human oversight.

DOI: https://doi.org/10.5281/zenodo.24114810