大语言模型发展综述  

Overview of the Development of Large Language Models

作　　者：闫啸彤 唐晓彬 沈童 雷诺 YAN Xiao-tong;TANG Xiao-bin;SHEN Tong;LEI Nuo(School of Statistics,University of International Business and Economics,Beijing 100029;Data Science Department,New York University,NY 10011,USA)

机构地区：[1]对外经济贸易大学统计学院,北京100029 [2]美国纽约大学数据科学系,纽约10011

出　　处：《统计学报》2024年第4期13-18,共6页Journal of Statistics

基　　金：国家社会科学基金重点项目(21ATJ001)。

摘　　要：在分析大语言模型特征的基础上,阐述了大语言模型的技术进展与应用领域,认为大语言模型在模型设计、数据结构、预训练和微调方面进步显著,展示了卓越的语言生成、知识运用和复杂逻辑推理等特性,这些特性使其在构建专业模型、提升知识整合和梳理水平、数据挖掘和分析、智能设计与操作领域展现出极大的应用潜力和实用价值。然而,目前模型在可解释性、安全性、真实性及训练成本等方面仍存在诸多挑战。随着算法和结构的优化,模型将实现更高效、更精确的训练,文本学习与生成性能将进一步增强,应用范围将拓展到多模态领域,实现多种类型数据的综合理解与处理。This paper analyzed the characteristics of large language models and elucidated their technological advancements and application domains.It was posited that large language models have made significant strides in model design,data architecture,pre-training and fine-tuning,exhibiting exceptional capabilities in language generation,knowledge application,and complex logical reasoning.These attributes enabled them to demonstrate immense potential and practical value in constructing specialized models,enhancing knowledge integration and parsing capabilities,data mining and analysis,as well as intelligent design and operation.Nonetheless,current models still faced numerous challenges regarding explainability,security,authenticity,and training costs.As algorithms and structures have optimized,these models were anticipated to achieve more efficient and precise training,further enhancing text learning and generation capabilities.Their application scope was set to expand into multi-modal realms,facilitating comprehensive understanding and processing of diverse data types.

关 键 词：大语言模型 语言生成 知识运用 复杂逻辑推理 数据挖掘 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]