Xiangyu Liu

Xiangyu Liu is a PhD student and Affiliate Assistant Professor at the University of Maryland. Their research focuses on advancing reinforcement learning (RL) and game theory, particularly in adversarial settings, multi-agent systems, and theoretical guarantees for partially observable environments. Advisor: Kaiqing Zhang.

Research interests include robust RL under adversarial perturbations, game-theoretic approaches to RL, LLM alignment challenges, and security aspects of AI systems. Recent work explores data poisoning threats to LLM alignment, regret analysis in online learning agents, and information-sharing strategies in multi-agent RL.

Publications emphasize theoretical foundations and practical defenses against adversarial attacks, with contributions to partially observable MDPs and non-dominated policy frameworks. Current research trends show strong focus on unifying behavioral diversity in zero-sum games and developing adaptive defense mechanisms against temporally-coupled perturbations.

No scientific awards explicitly listed in the provided text. Advising and grants: Currently advising no students (as a PhD student). Lab affiliation likely part of Kaiqing Zhang's research group, focusing on autonomous decision-making and AI safety.