Speaker: Jia Li Peking University
Time: 2024-12-23 10:00-2024-12-23 11:00
Venue: FIT 1-222

Abstract:

Software is the soul of next-generation information technology, and the ability to efficiently build software systems has become a core competitive advantage for nations. Intelligent software development aims to leverage artificial intelligence technologies to improve both the efficiency and quality of software development. My doctoral research focuses on automatically generating source code satisfying natural language requirements, known as code generation. In this talk, I will share our recent contributions to code generation with large language models (LLMs). (1) I will introduce aiXcoder-7B, an LLM for code. It contains multiple technical innovations (e.g., structured code infilling) and outperforms other LLMs with similar sizes, such as Meta's Code Llama-7B. (2) I will present ReasonCoder, an LLM-based code generation approach. It employs a novel multi-step reasoning paradigm and outperforms 25 other approaches across 8 benchmarks covering 5 programming languages, inspiring follow-up work. (3) I will describe EvoCodeBench, the first evolving code generation benchmark aligned with real-world repositories. It reveals the programming capabilities of LLMs in real software development. Our proposed code generation models and benchmarks have been successfully applied in many top IT companies, serving tens of thousands of developers.

Short Bio:

Jia Li is a final-year doctoral student at Peking University. His supervisors are Professor Zhi Jin and Professor Ge Li. His research direction is intelligent software development, specifically using artificial intelligence technologies to improve the efficiency and quality of software development. His doctoral studies focus on code generation and systematically contribute to the field. To date, he has published 15 papers in CCF-recommended A-level venues, 9 of which as the first author, including ICSE, ASE, FSE, NeurIPS, ACL, and AAAI. Researchers from MIT and Stanford University have frequently cited his papers. His studies have been successfully applied in many top IT companies, serving tens of thousands of developers. He has received the Excellent PhD Student Award at ChinaSoft 2023, the National Scholarship, and the Academic Top 10 Award from the School of Computer Science at Peking University.