Xiangyu Liu

xiangyu.jpg

2104 Brendan Iribe Center

College Park, Maryland 20740

I am a fifth-year PhD student in Computer Science at the University of Maryland, College Park (UMD) working under the guidance of Prof. Kaiqing Zhang since 2021, and an incoming research scientist at Google Research, NYC. I have also had the privilege of working with Prof. Furong Huang. Prior to this, I completed my undergraduate studies in Computer Science at Shanghai Jiao Tong University (SJTU) from 2017 to 2021, where I conducted my bachelor thesis research with Prof. Ying Wen.

My research focuses on the foundational aspects of (multi-agent) reinforcement learning (RL), particularly on game-theoretical/strategic (NeurIPS 2021, ICML 2023) and partially observable settings (NeurIPS 2024), as well as their applications on adversarially robust RL (ICLR 2024).

More recently, I have expanded my research to explore (multi-)large language model (LLM) agents interactions (ICLR 2025, ICML 2026), leveraging techniques from game theory and online learning.

In the past, I completed research internships at Google Research, Market Algorithms Team, where I worked with Zhe Feng, Aranyak Mehta, and Di Wang from Jun. 2025 to Dec. 2025, as well as Bloomberg AI Research in the summer of 2022.

News

Apr 30, 2026 My internship paper at Google Research is accepted to ICML 2026 and available on arxiv! We have explored scaling inference-time computation via opponent simulation for LLM-based strategic reasoning.
Apr 18, 2025 Gave an invited talk at UVA RL Meetup.
Jan 22, 2025 One paper accepted to ICLR 2025 on LLM agents.
Sep 25, 2024 One paper accepted to NeurIPS 2024 on partially observable RL with privileged information.
Mar 24, 2024 Gave a talk at 2024 INFORMS Optimization Society Conference (IOS 2024) on partially obseravble multi-agent RL at Houston, Texas.
Jan 16, 2024 Three papers accepted to ICLR 2024 including one spotlight.

Selected publications

* denotes equal contribution, † denotes alphabetical order.

  1. NeurIPS 2021
    Towards Unifying Behavioral and Response Diversity for Open-ended Learning in Zero-sum Games
    Xiangyu Liu, Hangtian Jia, Ying Wen, Yujing Hu, Yingfeng Chen, Changjie Fan, Zhipeng Hu, and Yaodong Yang
    Advances in Neural Information Processing Systems, 2021
    Code Website
  2. ICML 2023
    Partially Observable Multi-agent RL with (Quasi-)Efficiency: The Blessing of Information Sharing
    Xiangyu Liu, and Kaiqing Zhang
    In International Conference on Machine Learning, 2023
    Extended version accepted to SIAM Journal on Control and Optimization (SICON)
    arXiv
  3. NeurIPS 2024
    Provable Partially Observable Reinforcement Learning with Privileged Information
    Yang Cai†, Xiangyu Liu†, Argyris Oikonomou†, and Kaiqing Zhang†
    In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024
    A shorter version also presented at ICML 2024 ARLET workshop.
    arXiv
  4. ICLR 2024
    Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL
    Xiangyu Liu, Souradip Chakraborty, Yanchao Sun, and Furong Huang
    In The Twelfth International Conference on Learning Representations, 2024
    A preliminary version is selected at NeurIPS 2022 Trustworthy and Socially Responsible Machine Learning (TSRML) workshop as an outstanding paper
    arXiv Code
  5. ICLR 2024 (spotlight)
    Beyond Worst-case Attacks: Robust RL with Adaptive Defense via Non-dominated Policies
    Xiangyu Liu*, Chenghao Deng*, Yanchao Sun, Yongyuan Liang, and Furong Huang
    In The Twelfth International Conference on Learning Representations, 2024
    arXiv Code Website
  6. ICLR 2025
    Do LLM Agents Have Regret? A Case Study in Online Learning and Games
    Chanwoo Park*, Xiangyu Liu*, Asuman Ozdaglar, and Kaiqing Zhang
    In The Thirteenth International Conference on Learning Representations, 2025
    arXiv
  7. CDC 2025
    Principled Learning-to-Communicate with Quasi-Classical Information Structures
    Xiangyu Liu†, Haoyi You†, and Kaiqing Zhang†
    In 2025 IEEE 64th Conference on Decision and Control (CDC), 2025
  8. ICML 2026
    Scaling Inference-Time Computation via Opponent Simulation: Enabling Online Strategic Adaptation in Repeated Negotiation
    Xiangyu Liu, Di Wang, Zhe Feng, and Aranyak Mehta
    In International Conference on Machine Learning, 2026
    arXiv

Talks

Awards

Service

  • Reviewer for NeurIPS 2024-2025, ICLR 2025-2026, ICML 2025, AISTATS 2025-2026, UAI 2024-2025, AAMAS 2025-2026, CDC 2025, L4DC 2026
  • Graduate lecturer at summer AI camps for K12 students (2023, 2024), teaching multi-agent RL.