大模型:基于自然交互的人机协同软件开发与演化工具带来的挑战  

Challenges from LLMs as a Natural Language Based Human-machine Collaborative Tool for Software Development and Evolution

作　　者：李戈 彭鑫[2] 王千祥 谢涛 金芝 王戟[4] 马晓星[5] 李宣东[5] LI Ge;PENG Xin;WANG Qian-Xiang;XIE Tao;JIN Zhi;WANG Ji;MA Xiao-Xing;LI Xuan-Dong(School of Computer Science,Peking University,Beijing 100871,China;School of Computer Science,Fudan University,Shanghai 200438,China;Huawei Technologies Co.Ltd.,Beijing 100095,China;College of Computer Science and Technology,National University of Defense Technology,Changsha 410073,China;Department of Computer Science and Technology,Nanjing University,Nanjing 210023,China)

机构地区：[1]北京大学计算机学院,北京100871 [2]复旦大学计算机科学技术学院,上海200438 [3]华为技术有限公司,北京100095 [4]国防科技大学计算机学院,湖南长沙410073 [5]南京大学计算机科学与技术系,江苏南京210023

出　　处：《软件学报》2023年第10期4601-4606,共6页Journal of Software

摘　　要：以自然语言生成为核心的大模型技术正在人工智能领域掀起热潮,并持续向更多的领域穿透其影响力.以ChatGPT为代表的自然语言生成大模型(以下简称大模型),已经在软件工程的多项活动中展示出其通过自然交互方式给人提供一定程度帮助的能力和潜力,正在发展成为一种基于自然交互的人机协同软件开发与演化工具.从人机协同软件开发与演化的视角,大模型作为一种软件工具呈现出了两大特征:其一是基于自然语言的人机交互,在相当大程度上拓展了人机协同的工作空间、提高了人机协同的效率和灵活性;其二是基于已积累的软件开发和演化知识、针对给定软件开发和演化任务的预测性内容生成,可以对软件开发和演化工作提供一定程度的支持和帮助.然而,由于大模型本质是基于概率与统计原理和训练数据所形成的数学模型,具有不可解释性和内生不确定性,其生成的是缺失可信性判断的预测性内容,而人在软件开发与演化中所需要完成的是具有可信保障的决策性任务,所以大模型作为一种软件工具,在人机协同的软件开发和演化工作环境中给人提供帮助的同时,也带来了诸多的挑战.围绕如何构造对软件开发与演化更有帮助的代码大模型、如何引导大模型生成对软件开发与演化更有帮助的预测性内容、如何基于大模型生成的预测性内容开发与演化高质量的软件系统等大模型带来的挑战进行分析和阐述.The generative pertained transformer-based large language models(LLMs)are setting off a wave in the field of artificial intelligence and continue to penetrate their influence into more fields.The LLMs such as ChatGPT have demonstrated their ability and potential to provide people with a certain degree of assistance through natural language-based interaction in many software engineering tasks,and they are developing into a natural language-based human-machine collaborative tool for software development and evolution.From the perspective of human-machine collaborative software development and evolution,the LLMs,as a software tool,present two major features.One is the natural language-based human-machine interaction,which greatly expands the human-machine collaboration workspace and improves the efficiency and flexibility of human-machine collaboration.The second is to generate predictive contents based on accumulated knowledge of software development and evolution,targeting a given software development and evolution task,which can provide a certain degree of support and assistance for the software development and evolution task.However,since LLMs are essentially mathematical models based on probability and statistical principles and training date,with inexplicability and uncertainty,the contents generated by LLMs are predictive and lack the judgments for trustworthiness.As opposed to the tasks that humans need to perform in software development and evolution,which are typically decision-making tasks with trustworthiness guarantees,LLMs,as a software tool,not only provide assistance to people in software development and evolution featuring human-machine collaboration but also bring many challenges.This study analyzes and clarifies the challenges brought by the LLMs,such as how to construct LLMs that are more helpful for software development and evolution,how to guide LLMs to generate predictive contents that are more helpful for software development and evolution,and how to develop and evolve high-quality software s

关 键 词：软件开发与演化 大语言模型 人机协同 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]