Haoyue Bai's homepage

Haoyue BAI

Hi, my name is Haoyue Bai. I am a Ph.D. candidate at the Computer Sciences Department, University of Wisconsin-Madison, fortunately advised by Prof. Robert Nowak. I am a graduate visiting researcher in UC Berkeley, working with Prof. Dawn Song and Dr. Yiyou Sun. I am grateful to have worked with Prof. Sharon Li (UW-Madison), Dr. Wei Cheng (NEC Labs America), and Prof. Bolei Zhou (UCLA) during my graduate studies. I obtained my bachelor's degree at Zhejiang University and my master's degree at The Hong Kong University of Science and Technology, supervised by Prof. S.-H. Gary Chan.

My current research develops the theoretical and algorithmic foundations of reliable and trustworthy AI, with a focus on open-world robustness, data-efficient reliability, and safe reasoning in foundation models. A central theme of my work is leveraging post-deployment wild data, principled uncertainty, and provable objectives to ensure that AI systems can generalize, detect unknowns, and reason safely under distribution shift.

Out-of-distribution (OOD) learning and open-world robustness: Designing adaptive and interpretable learning principles that help ML models detect and generalize under distribution shifts, such as semantic and covariate shifts, correlation and diversity shifts.

Reliable algorithms with provable guarantees: Developing machine learning techniques with statistical and algorithmic guarantees to ensure reliable deployment of AI systems under real-world distribution shifts.

Safety and reliability of foundation models: Understanding the failure modes and boundaries of large language models (LLMs) and vision language models (VLMs) through systematic diagnostics, such as failure analysis in open-world reasoning, robustness under distribution shift, and generalization. Developing methods to strengthen their reliability and improve quality of learned representations and embeddings.

Data-efficient machine learning: Selecting the most informative data and signals (e.g., human or model feedback) to improve robustness, calibration, and coverage under tight annotation and compute budgets.

I am on the job market for 2025–2026. If you are interested in my research or background, please feel free to contact me.

Junior PhD/master/undergraduate students: If you would like to chat about life, career plan, graduate school applications, or research ideas related to AI/ML, feel free to email me to schedule a meeting. I will dedicate 30 mins every week, especially for students from underrepresented groups or whoever is in need.

Email: haoyue.bai [at] wisc [dot] edu

Google Scholar / Twitter / Linkedin

Recent Updates

Attending NeurIPS 2025 in San Diego from Dec 2rd to 7th.

One journal paper OoDBench+ accepted to TPAMI 2025.

Two papers accepted for presentation at NeurIPS 2025 MATH-AI.

Visiting Researcher at UC Berkeley with Prof. Dawn Song’s group this 2025 summer.

One poster and a lightning talk accepted to Agentic AI Summit 2025.

Passed my PhD preliminary exam and enter the dissertation stage in my third year at UW–Madison! Grateful to my committee: Prof. Rob Nowak, Prof. Yong Jae Lee, Prof. Fred Sala, and Prof. Tengyang Xie, for their guidance and support.

One paper on Black-box AI Content Detection accepted to CVPR 2025.

Two journal paper OLHF and ALOE accepted to TMLR 2025.

One paper AHA accepted to NeurIPS 2024.

I am a recipient of the EB-1A "Einstein Visa", conferred by the US Government for extraordinary acclaim in a specialized field.

Thanks OpenAI Superalignment Fellowship (worldwide 50 recipients) for supporting my research on frontier topics of foundation models!

One paper HYPO accepted to ICLR 2024.

One paper SCONE accepted to ICML 2023.

One Long Survey Paper accepted to Neurocomputing 2022.

One paper Quantifying and Understanding OOD Generalization accepted to CVPR 2022 as oral presentation.

Selected Publications

(See the full list on my Google Scholar page; ∗ indicates equal contribution.)

	RL Grokking Recipe: How Does RL Unlock and Transfer New Algorithms in LLMs? Yiyou Sun, Yuhan Cao, Bohao Huang, Haoyue Bai, Hannaneh Hajishirzi, Nouha Dziri, Dawn Song, Neural Information Processing Systems (NeurIPS) MATH-AI, 2025 paper / code / media coverage Can reinforcement learning (RL) actually teach large language models new algorithms or just “sharpen” what’s already latent in the base model? We set out to test this directly, and the finding is clear: RL can discover new capabilities, but only when trained wisely.
	Where’s the liability in the Generative Era? Recovery-based Black-Box Detection of AI-Generated Content Haoyue Bai, Yiyou Sun, Wei Cheng, Haifeng Chen, The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025 paper / code / media coverage We introduce a novel black-box detection framework that requires only API access, sidestepping the need for model weights or large auxiliary datasets.
	AHA: Adaptive Human-Assisted Out-of-Distribution Generalization and Detection Haoyue Bai, Jifan Zhang, Robert Nowak, Neural Information Processing Systems (NeurIPS), 2024 paper / code Our approach strategically labels examples within a novel maximum disambiguation region, where the number of semantic and covariate OOD data roughly equalizes. By labeling within this region, we can maximally disambiguate the two types of OOD data, thereby maximizing the utility of the fixed labeling budget.
	Provable Out-of-Distribution Generalization in Hypersphere Haoyue Bai, Yifei Ming, Julian Katz-Samuels, Yixuan Li International Conference on Learning Representations (ICLR), 2024 paper / code This paper provably learns domain-invariant representations in a hyperspherical space for OOD generalization.
	Feed Two Birds with One Scone: Exploiting Wild Data for Both Out-of-Distribution Generalization and Detection Haoyue Bai, Gregory Canal, Xuefeng Du, Jeongyeol Kwon, Robert Nowak, Yixuan Li International Conference on Machine Learning (ICML), 2023 Selected into UW-Madison CS762. [link] paper / code This paper bridges the gap between OOD generalization and OOD detection in one coherent framework.
	OoD-bench: Benchmarking and Understanding Out-of-Distribution Generalization Datasets and Algorithms Nanyang Ye, Kaican Li, Haoyue Bai, Runpeng Yu, Lanqing Hong, Fengwei Zhou, Zhenguo Li, Jun Zhu IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022 Selected as Oral Presentation. paper / code / media coverage This benchmark may serve as a strong foothold that can be resorted to by future OoD generalization research.
	NAS-OoD: Neural Architecture Search for Out-of-Distribution Generalization Haoyue Bai, Fengwei Zhou, Lanqing Hong, Nanyang Ye, S.-H. Gary Chan, Zhenguo Li IEEE International Conference on Computer Vision (ICCV), 2021 paper / code This work takes the first step to understand the OoD generalization of neural network architectures systematically.
	DecAug: Out-of-Distribution Generalization via Decomposed Feature Representation and Semantic Augmentation Haoyue Bai, Rui Sun, Lanqing Hong, Fengwei Zhou, Nanyang Ye, Han-Jia Ye, S.-H. Gary Chan, Zhenguo Li The AAAI Conference on Artificial Intelligence (AAAI), 2021 paper / code This work designs a general OoD generalization framework to tackle possible correlation shift and diversity shift in the real world.

Mentored Publications and Student Collaborations

( * indicates equal contribution; † indicates the mentoring role.)

	Deep Active Learning in the Open World Tian Xie, Jifan Zhang, Haoyue Bai † , Robert Nowak, Transactions on Machine Learning Research (TMLR), 2025 paper / code We introduce ALOE, a novel active learning algorithm for open-world environments designed to enhance model adaptation by incorporating new OOD classes via a two-stage approach.
	Unknown Aware AI-Generated Content Detection Ellie Thieu, Jifan Zhang, Haoyue Bai † under review, CVPR, 2025 We propose a learning framework for specific generator attribution that remains robust in the presence of unknowns or newly released generators.
	Towards Text-Guided Attribute-Disentangled Multimodal Representation Learning Yibing Wei, Sudeep Katakol, Manuel Brack, Jinhong Lin, Haoyue Bai †, Yu-Teng Li, Richard Zhang, Eli Shechtman, Hareesh Ravi, Ajinkya Kale under review, CVPR, 2025 This work identify a core limitation of current multimodal embeddings and formulate Queryable Attribute Representation Extraction (QARE) to explicitly evaluate query sensitivity and attribute invariance.
	CounTr: A Novel Transformer Approach for Image-based Crowd Counting Haoyue Bai † * , Hao He*, Zhuoxuan Peng, Tianyuan Dai, S.-H. Gary Chan European Conference on Computer Vision (ECCV) workshop, 2022 paper / code we introduce CounTr, a novel end-to-end transformer approach for crowd counting and density estimation, which enables capture global context in every layer of the Transformer.

Honors and Awards

OpenAI Superalignment Fellowship (worldwide 50 recipients across faculty and PhD students), 2024–present
OpenAI Researcher Access Program Recipient, 2024
CVPR Travel Support Award, 2025
Research Travel Grants, The Hong Kong University of Science and Technology, 2019
Outstanding Graduates, Zhejiang University, 2018
Alibaba–Zhejiang News Scholarship, 2017
Zhejiang University Scholarship & Awards: First-Prize Academic Scholarship (Top 3%), First-Class Scholarship for Outstanding Merits (Top 3%), Outstanding Social Worker Scholarship (Top 2%), Outstanding Student Leader Award (Top 6%), 2016–2017

Service

Department & Community
- ACM’s Women in Computing (WACM) Mentor at UW-Madison
- Graduate Student Mentor at UW-Madison
Seminar Organization
- Student Coordinator, Systems, Information, Learning and Optimization (SILO) Seminar
Professional Service
- Workshop Co-organizer: NeurIPS 2025 Workshop on Socially Responsible and Trustworthy Foundation Models (ResponsibleFM)
- Conference Reviewer: ICML, NeurIPS, ICLR; CVPR, ECCV, ICCV; AAAI, IJCAI, WACV
- Journal Reviewer: IEEE TPAMI, IEEE TNNLS, IEEE TIP

Teaching Assistant

University of Wisconsin–Madison
COMP CS540: Introduction to Artificial Intelligence

University of Wisconsin–Madison
COMP CS400: Data Science Programming III (Spring 2024)

University of Wisconsin–Madison
COMP CS220: Data Science Programming I (Fall 2022)

The Hong Kong University of Science and Technology
COMP2611: Computer Organization (Fall 2019, Spring 2019)

Other Research Experience

2024.5 -- 2024.8. NEC Laboratories America, Princeton.
Research Scientist Intern.
Mentors: Dr. Yiyou Sun, Dr. Wei Cheng.

2022.9 -- 2023.12. University of Wisconsin-Madison, Madison.
Research Assistant.
Advisor: Prof. Sharon Yixuan Li.

2021.6 -- 2021.9. The Chinese University of Hong Kong, Hong Kong.
Research Assistant.
Advisor: Prof. Bolei Zhou.

2018.1 -- 2018.6. The Hong Kong University of Science and Technology, Hong Kong.
Visiting Student.
Advisor: Prof. Shueng-Han Gary Chan.

2017.7 -- 2017.8. The Hong Kong University, Hong Kong.
Summer Intern.
Advisor: Prof. Lucas C.K. Hui.

How to Pronounce My Name

My name is Haoyue Bai. Haoyue is pronounced “How-yweh”. Bai is pronounced “bye”.

My first name 皓月 means “bright moon”: 皓 conveys brightness and whiteness; and 月 means "moon". My last name 白 means “white”. Together, my name evokes the image of a white, luminous moon shining in the night sky.

In Chinese culture, the moon symbolizes connection, shared emotions, and reunion across distance. A well-known line from Tang poetry says: “a bright moon shines across a thousand miles; we share this very moment.” I like to think of my name as carrying that same spirit — gentle light and warmth that bring best wishes to those we care about, no matter how far apart we may be.

Useful Links I Visit

How to read, write and present papers (by Prof. Vaidya at UIUC)

How to give a talk (by Profs. Jones, Hughes and Launchbury)

Student mentoring (by Prof. Wu)

Writing skills: Writing Technical Papers (from Prof. Jennifer Widom); Writing a math paper (by Profs. Kleiman and Tesled)

Task of a referee (good to know before write a paper)

Academic career: Advice from Professor David Patterson; Academic life

A math site

Thanks to Jon Barron for website template.