摘要
在人工智能技术快速发展的背景下,大模型技术作为推动科技进步和产业变革的重要力量,受到了学术界和产业界的广泛关注。作为一种大规模的预训练模型,大语言模型基于深度学习技术,拥有海量参数和强大的学习与泛化能力,能够处理和生成多种类型的数据。在回顾大语言模型技术发展历程的基础上,阐述了以Transformer架构为核心的基础架构设计,分析了当前通用的有监督微调、强化学习对齐等训练优化技术,探讨了大语言模型在应用与开发方面的实践。以DeepSeek系列模型为例,剖析了国产大语言模型的创新与应用。同时,简要阐述了大语言模型在实际应用中面临的挑战,如模型幻觉、效率瓶颈、价值对齐偏差等问题,并展望了未来大语言模型的发展趋势。
With the rapid development of AI,large-scale models have emerged as a driving force behind scientific and technological progress as well as industrial transformation,attracting widespread attention from both academia and industry.As large-scale pre-trained models,large language models(LLMs)based on deep learning technology possess an enormous number of parameters and exhibit powerful learning and generalization abilities,enabling them to process and generate various types of data.Based on a review of the development of LLMs,this paper discusses the architecture design centered around Transformer,analyzes current training and optimization techniques such as supervised fine-tuning and reinforcement learning alignment,and explores the practical applications and development of LLMs.Taking the DeepSeek series as an example,this paper examines the innovations and applications of domestic LLMs.Additionally,the challenges faced in real-world applications,such as model hallucinations,efficiency bottlenecks,and value alignment bias,are briefly discussed,and future development trends of LLMs are outlined.
作者
王文奇
郭梦帆
杨杜祥
张警元
洪飞阳
WANG Wenqi;GUO Mengfan;YANG Duxiang;ZHANG Jingyuan;HONG Feiyang(School of Computer,Zhongyuan University of Technology,Zhengzhou 450007,China)
出处
《中原工学院学报》
2025年第2期1-8,共8页
Journal of Zhongyuan University of Technology
基金
河南省重点研发专项(251111212000)。
作者简介
王文奇,男,教授,博士,主要研究方向为人工智能、网络安全、区块链技术.E-mail:5911@zut.edu.cn。
