About Me

I am a third-year PhD student in Cybersecurity at George Mason University, advised by Dr. Zhuangdi Zhu and expecting to graduate in 2028. My research focuses on LLM post-training, safety alignment, machine unlearning, and agentic AI.

One question runs through my work: once undesirable knowledge is in a model’s weights, what does it take to remove it — and to keep it removed? DUET (ICLR 2026) introduces on-policy distillation for unlearning, and CALIBURN (EMNLP 2026) reformulates it as policy-level preference optimization calibrated by the model’s own confidence. I also study the mirror image of the problem: keeping web content out of an LLM’s reach in the first place (EMNLP 2025).

Before Mason I completed my Master’s at the University of Chinese Academy of Sciences, working on privacy-preserving federated learning at the State Key Laboratory of Information Security.

In the summer of 2026 I interned on the AI team at The Washington Post, building LLM-based representations for personalized news recommendation.

I review for ICLR and IEEE Transactions on Information Forensics and Security. Reach me at yzhong7@gmu.edu — always happy to talk about unlearning, LLM safety, or collaborations.