Xiaoqian Lu

My research focuses on trustworthy machine learning and the safety and security of large language models (LLMs). In particular, I study vulnerabilities that can arise during LLM training and post-training, including data poisoning and backdoor attacks. My current work investigates how security risks can interact across different stages of LLM post-training, such as supervised fine-tuning and preference alignment, and explores ways to better understand and improve the robustness, reliability, and trustworthiness of modern AI systems.