Home
|
Yonggang Zhang
Research Assistant Professor @ HKUST
Address: Hong Kong University of Science and Technology
Clear Water Bay, Kowloon, Hong Kong
E-mail: zhangyg [at] ust.hk; yonggang9412 [at] gmail.com
[Google Scholar]
[GitHub]
|
Biography
My research focuses on trustworthy AI under deployment shift—the gap between
development-time assumptions and real-world requirements. I study how AI systems can
adapt to changing conditions, verify whether adaptations lead to genuine improvements,
and preserve existing capabilities. My work spans robust machine learning, reliable
adaptation of foundation models, and the self-evolution of autonomous agents.
My current research aims to establish the theoretical and algorithmic foundations of
reliable agent self-evolution, enabling autonomous agents to acquire reusable expertise
and improve through deployment experience while remaining reliable and accountable to
evolving human goals and permissions.
I am currently a Research Assistant Professor at the Hong Kong University of Science and
Technology (HKUST). Previously, I was a Postdoctoral Fellow at HKUST, working with
Prof. Yike Guo
(Fellow of IEEE and the Royal Academy of Engineering), and at Hong Kong Baptist
University (HKBU), working with
Prof. Yiu-ming Cheung
(IEEE Fellow) and
Prof. Bo Han.
I received my Ph.D. from the University of Science and Technology of China (USTC) in
2022, advised by
Prof. Xinmei Tian.
Research Interests
My research explores how AI systems can reliably adapt and improve under evolving
real-world conditions, with three interconnected directions:
-
Reliable Self-Evolution of Autonomous Agents:
learning from deployment experience, harness-level adaptation, verifiable
self-improvement, and governable agent systems.
-
Reliable Adaptation of Foundation Models:
targeted model adaptation, reasoning and alignment, and learning from reliable feedback.
-
Robust Machine Learning under Distribution Shift:
adversarial robustness, out-of-distribution generalization and detection,
and causal approaches to model reliability.
Selected Research
Selected work illustrating the development of my research from robust learning to reliable foundation models and autonomous agents. For a complete list, see the Publications page.
Verifiable Self-Evolution of Autonomous Agents
Reliable Adaptation of Foundation Models
-
Interpreting and Improving Large Language Models in Arithmetic Calculation (ICML 2024, Oral).
Identifying critical model components for targeted improvements in arithmetic reasoning.
-
Enhancing LLMs for Knowledge Base Question Answering by Chain-of-Decomposition (ICLR 2026).
Decomposing knowledge-base question answering to enable targeted adaptation.
-
Conditional Equivalence of DPO and RLHF: Implicit Assumption, Failure Modes, and Provable Alignment (ICML 2026, Spotlight).
Examining when preference optimization is aligned with intended preferences.
Robust Machine Learning
-
CausalAdv: Adversarial Robustness through the Lens of Causality (ICLR 2022).
Understanding adversarial vulnerability through causal data-generation mechanisms.
-
Detecting Generated Images by Fitting Natural Image Distributions (NeurIPS 2025, Spotlight).
Detecting generated images through properties of natural-image distributions.
Research Experience
-
Research Assistant Professor (Mar 2026 – Present)
Hong Kong University of Science and Technology (HKUST)
-
Postdoctoral Fellow (Jun 2025 – Feb 2026)
Hong Kong University of Science and Technology (HKUST)
Advisor: Prof. Yike Guo
-
Postdoctoral Fellow (Aug 2022 – Jun 2025)
Hong Kong Baptist University (HKBU)
Advisors: Prof. Yiu-ming Cheung
and Prof. Bo Han
-
Research Assistant (Dec 2020 – Jul 2022)
Hong Kong Baptist University (HKBU)
Advisor: Prof. Bo Han
Education
-
Ph.D. in Information and Communication Engineering (Sep 2017 – Jun 2022)
University of Science and Technology of China (USTC), China
Thesis: Adversarial Robustness of Deep Learning through the Lens of Data Distribution
Supervisor: Prof. Xinmei Tian
-
Bachelor of Electronic Information Engineering (Sep 2013 – Jun 2017)
China University of Petroleum (East China), Qingdao, China
Thesis: Research on Adversarial Examples
Supervisors: Yanjiang Wang and Weifeng Liu
|