Assistant Professor
AI Thrust, HKUST (GZ)
https://scholar.google.com/citations?hl=en&user=qb9STggAAAAJ
https://www.xiaohongshu.com/user/profile/63e87c990000000027028efa
:linkedin: Linkedin
Dr. Yao SHU (舒瑶) is now a tenure-track assistant professor of Artificial Intelligence Thrust at Hong Kong University of Science and Technology (Guangzhou), starting from April 8, 2025. He was previously a researcher at the Guangdong Lab of AI and Digital Economy (SZ) under the Genius Nova Programme. Before that, Dr. Yao SHU was a senior researcher (Tencent Technical Fellow) at Tencent in 2023 and a Research Fellow (PostDoc.) at the National University of Singapore (NUS) in 2022. He received his PhD from the NUS School of Computing in 2022 under the esteemed guidance of Prof. Bryan Low (Associate Vice President (AI) of NUS and Deputy Director of NUS AI Institute), where he was awarded the IMDA Excellence in Computing Prize for the best PhD thesis in the school, as well as the Dean's Graduate Research Excellence Award for significant research achievements during his doctoral studies. He earned his B.Eng. from Huazhong University of Science and Technology (HUST) in 2017, advised by Prof. Kun He (何琨). Dr. SHU's research focuses on developing advanced learning and optimization algorithms that are both theoretically sound and practically useful. He is publishing actively in top AI conferences, including ICML, NeurIPS, ICLR, AAAI, ACL, EMNLP, and UAI. He also serves as a reviewer for leading AI conferences such as ICML, NeurIPS, ICLR, AAAI, IJCAI, UAI, AISTATS, AAMAS, etc., and was recognized as a UAI 2025 Top Reviewer and ICLR 2025 Notable Reviewer. Furthermore, he is invited to serve as an Area Chair for ICLR.
Theory-First Optimization and Learning Research for Artificial Intelligence We aim to develop foundational principles and practically useful algorithms that reshape how AI systems are optimized and understood. By building unifying theoretical frameworks and translating them into more efficient, robust, and adaptive algorithms, we seek to advance the reliability of large-scale AI systems, especially large language models and intelligent agents.
1. Unified and Improved Black-Box Optimization We establish a coherent oracle framework covering ZOO, BO, bandits, and RL, and leverage it to design improved optimization algorithms that outperform classical techniques in efficiency and reliability.
2. Learning Theory and Algorithmic Principles We advance general learning theory to explain generalization, stability, robustness, and adaptation in complex environments. Guided by these insights, we develop algorithms that are provably sound and practically beneficial for large-scale learning under non-ideal conditions.
3. Algorithmic Impact via LLMs and Agents We validate and extend theoretical advances by applying them to LLMs and AI agents. This creates a feedback loop where theoretical principles inspire new algorithms, and real-world applications reveal new theoretical questions—ultimately improving efficiency, personalization, and safety in deployed AI.