Jian Yang
Associate Professor at the School of Computer Science, Beihang University
Jian Yang is an Associate Professor at Beihang University and a Huawei Distinguished Young Scholar. His research focuses on large code models and intelligent agents, with particular interests in pre-training, post-training, collaborative generation, and evaluation. He has published more than 100 papers in leading international journals and conferences, including ICLR, NeurIPS, ACL, and EMNLP, and has served as the first or corresponding author on more than 30 publications. He has also served multiple times as SAC/AC for ARR (ACL, EMNLP, and NAACL), Program Chair of INLG 2026, and SPC and AC for the AIA Track of AAAI. His work has received more than 40,000 citations on Google Scholar, and he was honored with the WACI 2026 Yunfan Award. After completing his Ph.D., he joined Alibaba’s Qwen team through the Alibaba Star talent program. As a core contributor to the Qwen series of large language models and the QwenCoder series of code models, he was responsible for advancing Qwen’s code capabilities and leading the development of specialized QwenCoder models. After joining Beihang University, he led the development of LoopCoder, a recurrent code foundation model, as well as InCoder-32B for industrial applications, covering multiple model scales including 7B, 14B, 32B, and 40B parameters. In the practical application and deployment of large code models, he has received the First Prize for Innovation and Entrepreneurship from the China Association of Inventions and holds more than ten patents, application certificates, and vulnerability validation certificates.
Topic
From Code Foundation Models to Repository-Level Agents: Core Technological Evolution and Industrial-Scale Deployment
This talk traces the evolution from code foundation models to code agents, highlighting key technological advances and research progress. It covers pre-training scaling laws for large code models, collaborative training across multiple programming languages, verifiable reinforcement learning, and multi-granularity, multilingual evaluation frameworks, as well as key capability enhancements for industrial applications, including controllable generation, multi-turn interaction, and factual question answering. The team will also provide an in-depth introduction to emerging directions such as repository-level software engineering agents, code execution and reflection, and multimodal development agents. The session will showcase the strong performance of its open-source models, including LoopCoder and Qwen-Coder, on authoritative benchmarks such as SWE-bench. By closely integrating theoretical innovation, large-scale data construction, and engineering practice, this presentation offers a systematic approach and open ecosystem support for building more reliable and intelligent code agents.
