Abstract: Source code is far more predictable and repetitive than natural language. And large language models of code have demonstrated their great potential for assisting various software tasks, such as ChatGPT (from OpenAI) for coding and software Q&A. However, are those existing AI models really good/practical enough to take over the programming job? This talk will answer the questions from 3 perspectives to show that the models are still facing a “trust crisis”, (1) the existing models are not effective under specific scenarios, (2) the existing models are too big to be deployed in a low-resource development environment, (3) the existing models cannot perceive code changes very well. Correspondingly, this talk will cover solutions to handle the 3 limitations. Besides, this talk will also present some ongoing works and future directions towards trustworthy code intelligence, such as from software data distribution perspective.

Short Bio: Bowen Xu is a Research Scientist (Post-Doc) in the School of Computing and Information Systems at Singapore Management University. His research area in general is software engineering. Particularly, he focuses on AI for code and NLP for SE. The aim of his work is to improve software quality and developers’ productivity. He has published and presented 22 research works in top-tier software engineering conferences (ICSE, FSE, ASE) and journals (TSE, EMSE). One of his recent works has won the Honorable Mention Award for ACSAC (Core A). Another recent work from him is nominated for ACM SIGSOFT Distinguished Paper Award (top 10% among all accepted submissions) for ASE 2022 (Core A*). Besides, his work published in ESEM (Core A) has won the Highly Commended Full Paper Award (the 2nd best among all accepted submissions). Two of his works published in ICPC (Core A) and MSR (Core A) have been invited extended to the journal EMSE.