大模型代码生成技术及航天领域潜在应用  

Large model code generation technology and its potential applications to aerospace

作　　者：陈晓阳 高飞[1] 韩翔宇[1] 马卫华 CHEN Xiaoyang;GAO Fei;HAN Xiangyu;MA Weihua(Beijing Aerospace Automatic Control Institute,Beijing 100854,China;Beijing Shenzhou Aerospace Software Technology Co.,Ltd,Beijing 100094,China)

机构地区：[1]北京航天自动控制研究所,北京100854 [2]北京神舟航天软件技术股份有限公司,北京100094

出　　处：《航天控制》2025年第1期8-16,共9页Aerospace Control

摘　　要：考虑到基于大语言模型(LLMs)的代码生成技术对软件生产力的巨大影响及在航天领域的应用前景广泛,本文从问题背景与定义、典型技术与其在航天领域的潜在应用场景以及应用评价方法3个方面,综述了该技术的最新研究进展,以期为航天领域代码生成技术的相关研究提供指导与启发。首先,从代码生成问题定义及LLMs的结构特点,讨论了LLMs在代码生成方面的基础能力;然后,在此基础上,详述了包括预训练技术、指令微调技术、提示词工程和检索增强技术等实现代码生成的主要方法及其在航天领域的潜在应用场景;接着,从语义相似性和验证数据集两方面,梳理了评估基于LLMs的代码生成技术的主要方法,并分析了它们的特点及局限性;最后讨论了LLMs技术在代码生成问题中所面临的挑战及未来发展方向。Regarding the significant impact of code generation techniques based on large language models(LLMs)on software productivity and their broad application prospects to the aerospace field,the latest research progress on this kind of technology is reviewed from three aspects:problem background and definition,typical technologies and their potential application scenarios in the aerospace domain,and application evaluation methods,with the aim of providing guidance and insights for related research on code generation techniques in the aerospace domain.Firstly,the basic capabilities of LLMs are dis⁃cussed on code generation according to the features of code generation problem definition and LLMs struc⁃tures.Then,the main methods for code generation are elaborated,including pre-training,instruction fine-tuning,prompt engineering and retrieval-augmented generation as well as their potential application scenarios to the aerospace field.Next,due to the perspectives of semantic similarity and validation datas⁃ets,the popular methods are reviewed for evaluating the results of LLM-based code generation tech⁃niques and their characteristics and limitations are analyzed.Finally,the challenges are presented and future improvements are proposed.

关 键 词：代码生成 大语言模型 预训练 指令微调 检索增强 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]