Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models
Reasoning models hide unsafe thoughts behind safe answers; we propose a metric DSAR to measure it and safety alignment method SARA to mitigate it.
About Me
I am a Ph.D. candidate in Computer Science at Wayne State University, advised by Prof. Dongxiao Zhu in the Trustworthy AI Lab. My research focuses on trustworthy AI: making AI systems robust and aligned so they can be deployed safely.
I work across large language models (LLMs), large reasoning models (LRMs), and agentic AI, spanning safety alignment, fine-tuning, reinforcement learning, in-context learning, and machine unlearning. My work has been published at venues including NeurIPS, ICLR, CVPR, and AAAI (oral). In industry, I worked on generative retrieval as an AI/ML engineer intern at LinkedIn.
I am always happy to connect with others working on trustworthy AI. Feel free to reach out!
Education
Ph.D. in Computer Science
2023-08-30
Wayne State University
M.S. in Computer Science
2021-08-30
2023-05-30
Stevens Institute of Technology
B.S. in Software Engineering
2017-09-01
2021-05-30
Chongqing University of Posts and Telecommunications
Interests
Trustworthy AI
Large Language Models
Large Reasoning Models
Generative Retrieval
Agentic AI
Featured Publications
Hijacking Large Language Models via Adversarial In-Context Learning
This work introduces a novel transferable attack against In-Context-Learning to hijack LLMs to generate the target response or jailbreak. We also propose a defense strategy …
Not all tokens are meant to be forgotten
This paper helps large language models forget sensitive and unwanted data without over-forgetting general data.
Recent Publications
(2026).
Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models.
NeurIPS-26.
(2026).
Hijacking Large Language Models via Adversarial In-Context Learning.
ECML-PKDD-26.
(2026).
Not all tokens are meant to be forgotten.
AAAI-26.
(2026).
WalkGPT Grounded Vision-Language Conversation with Depth-Aware Segmentation for Pedestrian Navigation.
CVPR-26.
(2026).
Attention Smoothing Is All You Need For Unlearning.
ICLR-26.
Recent & Upcoming Talks
Not All Tokens Are Meant to Be Forgotten
Oral presentation of our AAAI-26 paper on Targeted Information Forgetting (TIF), a framework that unlearns unwanted information at the token level without collapsing model utility.
Recent News
🎉 Paper accepted by NeurIPS-26
Our paper 'Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models' has been accepted to the Fortieth Annual Conference on Neural Information Processing Systems …
💼 Summer 2026 at LinkedIn as an AI/ML Engineer Intern
I spent summer 2026 on LinkedIn's Generative AI team in Mountain View, building the group's first reinforcement-learning pipeline for generative candidate retrieval in the hiring …
Will be serving as one of the reviewer at ICML 2026
I was glad to serve as a reviewer for ICML 2026 this year. Seeing papers from the reviewer side gave me a better sense of what makes research stand out.
🎉Paper accepted by AAAI-26
Excited to share our paper 'Not All Tokens Are Meant to Be Forgotten' has been accepted to the The 40th Annual AAAI Conference on Artificial Intelligence (Accepted as Oral, …
Serving as one of the Program Committee (PC) members at AAAI 2026
I will be serving as one of the Program Committee (PC) members at AAAI 2026.
