Large Language Models and Their Evolution

Abstract

With the rapid advancement of artificial intelligence (AI), large language models (LLMs) have become the foundational infrastructure for natural language processing (NLP) research and industrial applications. By leveraging massive parameters and vast pre-training data, LLMs have significantly enhanced text understanding, generation, and cross-modal reasoning capabilities. This paper systematically reviews the technical evolution of LLMs from n-gram statistical models to the Transformer architecture, based on five key review papers. It analyzes training and alignment paradigms such as pre-training & fine-tuning and RLHF/RLAIF, as well as the exponential parameter expansion driven by the Scaling Law. Furthermore, we summarize the latest application progress of LLMs in code generation, intelligent customer service, medical and legal assistance, and other fields, and analyze the challenges they face in terms of data privacy, model bias, hallucination phenomena, and energy consumption. Finally, this paper proposes four research priorities for the future: first, leveraging explainable mechanisms to enhance model transparency; second, strengthening value alignment and security controls; third, exploring green and efficient model compression and inference schemes; and fourth, leveraging interdisciplinary collaboration to build the next generation of general-purpose intelligent systems that are both fair and sustainable.