Xu Hu
☰

I have been a Ph.D. student in Computer Science at the University of Texas at Dallas (UT Dallas) since fall 2025, under the guidance of Prof. Feng Chen (Fellow, Eugene McDermott Distinguished Professor) and Prof. Bingzhe Li (Assistant Professor).

My research focuses on computer-use agents (CUA): agents that automate long, multi-application computer work. I also work on persona agents as a member of the MatrAIx organization. My current research agenda covers four directions:

Direction Representative Work Core Question
Agent post-training

Poly-CUA (in progress): multi-agent reinforcement learning for computer-use agents.

LEAD (NeurIPS 2026): an RL method that curbs overly long chain-of-thought reasoning in LLMs.

 How can post-training make agents more capable on long-horizon computer tasks, and their reasoning more efficient?
Benchmarks for CUA

CUA-Recovery (under review at ICLR): a long-horizon robustness benchmark for computer-use agents.

PersonaCUABench (under review at NAACL): the first benchmark to evaluate how persona influences agent task performance.

 How robust are computer-use agents over long horizons, and how much does the user's persona change task performance?
LLM reliability

Small Updates, Big Doubts (TMLR 2026): an empirical study of how parameter-efficient fine-tuning affects hallucination detection.

Toward Trustworthy Computer-Use Agents (under review at ACM Computing Surveys): a survey of risk propagation, evaluation gaps, and human governance in computer-use agents.

 Does parameter-efficient fine-tuning make hallucinations easier to detect, and where do risks arise and propagate when agents operate computers?
Agent memory

SAFE-Mem (under review): a secure memory framework that labels every stored claim with its kind, authority, and scope, and retrieves it only for the uses its source permits.

Anatomy of Agentic Memory (under review): a taxonomy and empirical analysis of agentic memory systems and how they are evaluated.

 Where do current agentic memory systems, and the ways we evaluate them, fall short, and how can memory block untrusted content without losing legitimate information?

Education

  • 2025–present University of Texas at Dallas (Richardson, TX, USA) — Ph.D. in Computer Science.
  • 2021–2025 Anhui University (Hefei, China) — B.S. in Statistics (ranked 1st of 103 in my major).

Research Experience

  • 2024 North Carolina State University (Raleigh, NC, USA) — visiting student in Statistics and Operations Research, with Prof. Xiaolei Fang (Associate Professor) and Prof. Shu-Cherng Fang (Walter Clark Chair Professor and University Alumni Distinguished Graduate Professor; Fellow of the Institute of Industrial Engineers).
  • 2024 Georgia Tech & Emory University — remote research intern in Biomedical Engineering at the TReNDS Lab with Prof. Vince D. Calhoun (Professor of Biomedical Engineering, Georgia Tech & Emory; Distinguished University Professor, Georgia State University).
  • 2024 HKUST (Guangzhou), Information Hub — research assistant with Yuxuan Liang and Zhengyang Zhou.
  • 2023 Nanyang Technological University (Singapore) — AI Summer School; received the Best Project Award and Best Leadership Award (sole recipient).

Honors & Awards

  • 2025 National Scholarship (China).
  • 2023 Best Project Award and Best Leadership Award (sole recipient), Nanyang Technological University AI Summer School.
  • 2023 Finalist (Top 1.5%), Interdisciplinary Contest in Modeling (ICM).

News

  • Oct 2026 New preprint: CUA-Recovery — a test suite of 2,250 human-verified erroneous states and a training suite of 8.6K recovery trajectories for long-horizon computer-use agents; across six agents, error awareness peaks at 44.0%. Code is available.
  • Oct 2026 New preprint: PersonaCUABench — the same 74 computer-use tasks requested by 44 personas in their own words; task completion drops by 36% compared with the original instructions. Project page and code are available.
  • Oct 2026 New preprint: SAFE-Mem — secure agentic memory with fine-grained authority and governance-aware retrieval. Code is available.
  • Oct 2026 Happy to contribute to and share this interesting work: PersonalAgentEval, a leaderboard for evaluating personal agents. More explorations of persona agents are coming soon. Stay tuned!
  • Sep 2026 No Hidden Prompts Needed! was accepted to the NeurIPS 2026 Workshop on AI-Native Academia. See you in Atlanta in December!
  • Sep 2026 LEAD was accepted to NeurIPS 2026.
  • Sep 2026 New work: CUA-Recovery — a benchmark and training set for error recovery in long-horizon computer-use agents.
  • Jul 30, 2026 Our paper Small Updates, Big Doubts, on hallucination detection, was accepted for publication in Transactions on Machine Learning Research (TMLR). This is the first paper from my Ph.D. research!
  • Jun 2026 New preprint: No Hidden Prompts Needed! — gaming AI peer review with presentation-only revisions.
  • May 2026 New preprint: LEAD — length-efficient adaptive and dynamic reasoning for LLMs.
  • 2026 New survey: Toward Trustworthy Computer-Use Agents — risk propagation, evaluation gaps, and human governance. Now under review at ACM Computing Surveys.
  • Feb 2026 Two new preprints: Small Updates, Big Doubts (hallucination detection) and Anatomy of Agentic Memory.
  • Oct 2025 Paper at IEEE MASS 2025: Can Large Language Models Be Trusted Paper Reviewers?
  • Sep 2025 GM-LDM accepted to the PRIME 2025 workshop at MICCAI 2025.
  • Aug 2025 Started my Ph.D. in Computer Science at UT Dallas.

Publications

Bold indicates my name; † denotes equal contribution. For the latest list, see my Google Scholar.

  1. Overview of PersonaCUABench: who asks, interaction, and scoring
    PersonaCUABench: Benchmarking Computer-Use Agents Serving Diverse Personas. under review
    Haoming Li†, Defu Cao†, Xu Hu†, Yingtao Luo, Yiyang Wang, Xianghe Xu, Jiashu He, Wanting Jiang, Yifan Zhang, Yunhan Zhao, Siyang Wu, Chen Yang, Haixin Wang, Zheyuan Deng, Taoran Li, Jingwen Zhang, Zhejian Peng, Shushu Liang, Yue Zhao, Bingzhe Li, Feng Chen, Yifan Wang, Jintao Huang, Yijun Wang, Yuexing Hao, Xiaomin Li.
    arXiv preprint arXiv:2610.22990, 2026.
    arXiv project page code
  2. Overview of SAFE-Mem: governance labels, memory graph, and governance-aware retrieval
    SAFE-Mem: Secure Agentic Memory with Fine-Grained Authority and Governance-Aware Retrieval. under review
    Haoming Li, Dingyi Kang, Xu Hu, Jinxin Yang, Zhu Xu, Chenru Wang, Yi Ge, Miao Yin, Guanpeng Li, Feng Chen, Bingzhe Li.
    arXiv preprint arXiv:2610.22959, 2026.
    arXiv code
  3. Overview of the CUA-Recovery benchmark and its construction
    CUA-Recovery: Benchmarking Error Recovery in Long-Horizon Computer-Use Agents. preprint
    Xu Hu†, Haoming Li†, Yuxiang Wang, Jinxin Yang, Kangshuo Li, Licheng Wang, Dingyi Kang, Jiayu Song, Tianhao Wang, Zhanhua Hu, Xu Yang, Qiannan Li, Bingzhe Li, Feng Chen.
    arXiv preprint arXiv:2610.27645, 2026.
    arXiv code
  4. Overview of the presentation-only review-gaming framework
    No Hidden Prompts Needed! You Can Game AI Peer Review with Presentation-Only Revisions. NeurIPS 2026 Workshop
    Xu Yang, Zhizhou Sha, Junbo Li, Jian Yu, Yifan Sun, Matthew Zhao, Jinrui Fang, Xinyue Guo, Yining Wu, Xu Hu, Yifu Luo, Qiang Liu, Zhangyang Wang.
    NeurIPS 2026 Workshop on AI-Native Academia: Authorship, Peer Review, and Conference Governance under AI, 2026.
    arXiv project page code
  5. The LEAD framework
    LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models. NeurIPS 2026
    Songtao Wei, Yi Li, Zhikai Li, Xu Hu, Yuede Ji, Guanpeng Li, Feng Chen, Carl Yang, Zhichun Guo, Bingzhe Li.
    Advances in Neural Information Processing Systems (NeurIPS), 2026.
    arXiv code
  6. Overview of risk propagation in computer-use agents
    Toward Trustworthy Computer-Use Agents: Risk Propagation, Evaluation Gaps, and Human Governance. under review
    Xu Hu, Haoming Li, Sabbir Ahmed, Md Nahiyan Uddin, Qiannan Li, Jessica Ouyang, Latifur Khan, Feng Chen, Bingzhe Li.
    Under review at ACM Computing Surveys, 2026.
    preprint paper list
  7. Taxonomy of memory-augmented generation systems
    Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitations. under review
    Dongming Jiang, Yi Li, Songtao Wei, Jinxin Yang, Ayushi Kishore, Alysa Zhao, Dingyi Kang, Xu Hu, Feng Chen, Qiannan Li, Bingzhe Li.
    arXiv preprint arXiv:2602.19320, 2026.
    arXiv paper list
  8. Overview of the empirical study of fine-tuning and hallucination detection
    Small Updates, Big Doubts: Does Parameter-Efficient Fine-tuning Enhance Hallucination Detection? TMLR
    Xu Hu, Yifan Zhang, Songtao Wei, Chen Zhao, Qiannan Li, Bingzhe Li, Feng Chen.
    Transactions on Machine Learning Research (TMLR), 2026.
    arXiv
  9. Agent-based paper review and interactive discussion pipeline
    Can Large Language Models Be Trusted Paper Reviewers? A Feasibility Study. IEEE MASS 2025
    Chuanlei Li, Xu Hu, Minghui Xu, Kun Li, Yue Zhang, Xiuzhen Cheng.
    2025 IEEE 22nd International Conference on Mobile Ad-Hoc and Smart Systems (MASS), pp. 531–536, IEEE, 2025.
    paper
  10. Architecture of the GM-LDM 3D autoencoder and conditional denoising network
    GM-LDM: Latent Diffusion Guided by Functional Connectivity for Gray-Matter Generation and Biomarker Identification. MICCAI 2025 Workshop
    Xu Hu, Jingling Yang, Wenjun Xiao, Sihan Jia, Yutong Gao, Zening Fu, Vince D. Calhoun, Yuda Bi.
    Predictive Intelligence in Medicine (PRIME 2025), Workshop at MICCAI 2025, pp. 58–69, 2025.
    paper code

Academic Service

  • Conferences Reviewer for ICLR 2027, NeurIPS 2026, and KDD 2027.
  • Journals Reviewer for the Journal of Data Science and Analysis and Frontiers.

Life & Interests

  • Background I started my undergraduate studies in law before moving to statistics. I also hold the ACCA (Association of Chartered Certified Accountants) certificate.
  • Interests Outside research, I am interested in business, law, sociology, and psychology.