王翀
内容简介
题目
Leveraging Code LLMs and Static Analysis for Practical Software Development
摘要
Recent code large language models (LLMs) have shown promising performance in solving software development tasks such as type inference and code generation. However, they face limitations in application scenoies, due to the lack of project-specific contexts like user-defined types and functions. In this talk, I will first introduce TIGER, a two-stage generating-then-ranking framework, designed to effectively leverage pre-trained code models and static analysis to handle Python's diverse type categories, especially complex generic types and (unseen) user-defined types. Second, I will briefly introduce our recent work on integrating an autocompletion tool into LLM-based code generation to mitigate dependency issues such as no-member and undefined attribute errors in generated code. Finally, I will discuss future directions for enhancing practical software development by combining code LLMs with static analysis.
报告人
王翀,南洋理工大学博士后,于2023年和2018年分别从复旦大学获得博士学位和学士学位,导师为彭鑫教授。现在南洋理工大学任博士后研究员,合作导师为刘杨教授。研究方向为智能化软件工程,致力于应用大模型和知识图谱等前沿智能化技术解决软件开发中的实际问题。多项研究成果已发表于软件工程领域顶级会议与期刊ICSE、FSE、ASE、TSE、TOSEM等。曾获得CCF上海市优秀博士论文奖及IEEE杰出论文奖等奖项。
时间安排
❖
时间:2024年7月9日,10:30 – 12:00
❖
地点:复旦大学江湾校区交叉二号学科楼A2003