Xiaoyuan Zhu

Amos • 朱孝缘（Zhū Xiào-yuán）• xzhu9839 [at] usc [dot] edu

xiaoyuan.jpg

Amos • 朱孝缘（Zhū Xiào-yuán）• xzhu9839 [at] usc [dot] edu

Hi, I’m Xiaoyuan, an incoming PhD student at TTIC advised by Prof. Nicholas Tomlin. Previously, I was an undergrad at USCUSC logo, where I was fortunate to be advised by Prof. Willie Neiswanger, Prof. Robin Jia, and Prof. Xuezhe Ma. I had wonderful time interning at USC Interaction LabInteraction Lab logo (summer 2023), CAISCAIS logo (winter 2023), and with Prof. Steven Wu at CMU S3DCMU logo (summer 2025).

My research mainly focuses on developing efficient and effective training-time and inference-time methods to improve model safety and usability. Some broad questions I’m recently interested in:

  • Can we identify what can be learned from user feedback, and how to enable models to learn continuously after deployment?

  • How to improve the sampling and computational efficiency of rollouts in RL, especially in settings with sparser reward?

Alongside my research, I co-lead USC’s AI Safety Club, a student group promoting safety in advanced AI systems. We run semester-long programs on introductory and technical topics. I’ve also explored computational neuroscience 🧠.

news

Jan 26, 2026 Our paper 🔭 Hubble has been accepted to ICLR 2026.
Jan 26, 2026 Our paper 🔍 Auditing Black-Box LLM APIs has been accepted to ICLR 2026.
Jul 07, 2025 Our paper on 🧹 LLM Unlearning has been accepted to COLM 2025.
Oct 08, 2024 Our new version of 🦙 Analysis360 is released.
Apr 18, 2024 I received the Provost’s Undergrad Research Fellowship at USC.

publications

* Check my google scholar for a full list of publications

  1. Preprint
    Justified or Just Convincing? Error Verifiability as a Dimension of LLM Quality
    Xiaoyuan Zhu, Kimberly Le Truong, Riccardo Fogliato, Gokul Swamy,
    and 7 more authors
    arXiv 2604.04418, 2026
    arXiv
  2. ICLR
    Auditing Black-Box LLM APIs with a Rank-Based Uniformity Test
    Xiaoyuan Zhu, Yaowen Ye, Tianyi Qiu, Hanlin Zhu, and 5 more authors
    In The International Conference on Learning Representations, 2026
    arXiv
  3. COLM
    LLM Unlearning Without an Expert Curated Dataset
    Xiaoyuan Zhu, Muru Zhang, Ollie Liu, Robin Jia, and 1 more author
    In The Conference on Language Modeling, 2025
    arXiv Code Poster
  4. ICLR
    Hubble: a Model Suite to Advance the Study of LLM Memorization
    Johnny Tian-Zheng Wei*, Ameya Godbole*, Mohammad Aflah Khan*, Ryan Wang, and 6 more authors
    In The International Conference on Learning Representations, 2026
    arXiv Code Website
  5. NAACL
    Using Linguistic Entrainment to Evaluate Large Language Models for Use in Cognitive Behavioral Therapy
    Mina Kian, Kaleen Shrestha, Katrin Fischer, Xiaoyuan Zhu, and 6 more authors
    In the Findings of the Association for Computational Linguistics: NAACL, 2025
    PDF
  6. ICML
    The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning
    WMDP Team
    In The International Conference on Machine Learning, 2024
    arXiv Code Website
  7. AAAI-SS
    How Can Large Language Models Enable Better Socially Assistive Human-Robot Interaction: A Brief Survey
    Zhonghao Shi, Ellen Landrum, Amy O’Connell, Mina J. Kian, and 4 more authors
    In the AAAI Symposium Series, 2024
    arXiv
  8. EMNLP
    Learn Your Tokens: Word-Pooled Tokenization for Language Modeling
    Avijit Thawani, Saurabh Ghanekar, Xiaoyuan Zhu, and Jay Pujara
    In the Findings of The Conference on Empirical Methods in Natural Language Processing, 2023
    arXiv Code
  9. Interspeech
    Multimodal Speech Recognition for Language-Guided Embodied Agents
    Allen Chang, Xiaoyuan Zhu, Aarav Monga, Seoho Ahn, and 2 more authors
    In the Interspeech Conference, 2023
    arXiv Code