Yunbei Zhang

Yunbei Zhang

Ph.D. Candidate at Tulane · Currently at Oak Ridge National Lab · Previously at Amazon, KLA

My research aims to develop trustworthy, efficient, and adaptive AI systems, focusing on inference-time learning and reasoning in settings where neither ground-truth labels nor verified rewards are available, safety and red-teaming of frontier multimodal models, and efficient model adaptation for both Model-as-a-Service (MaaS) APIs and on-device deployment. I also work on post-training, reasoning, and planning for large language and vision-language models.

Inference-Time Learning AI Safety MLLM Reasoning & Planning MaaS & On-Device
Open to research internships and collaborations. If my work resonates with your interests, I'd be delighted to connect: yunbeizhang.ml [at] gmail.com

News

Feb '26 Two papers accepted to CVPR 2026 — see you in Denver!
Feb '26 Invited reviewer for ACM MM 2026
Feb '26 New preprint: Continual Test-Time Adaptation: A Comprehensive Survey
Feb '26 New preprint: Agents in the Wild studies safety and social dynamics on an AI-only platform
Feb '26 New preprint: Visual Exclusivity Attacks introduces a new threat model for multimodal red-teaming
Jan '26 Joined Oak Ridge National Laboratory as Research Scientist Intern
Jan '26 Survey on Prompt-based Adaptation in Vision Models accepted to TMLR
Jan '26 Invited reviewer for ECCV 2026 and ICML 2026
Jan '26 One paper accepted to ICASSP 2026
Dec '25 Received Lambda Research Grant
Oct '25 Awarded Top Reviewer at NeurIPS 2025
Oct '25 Invited reviewer for CVPR 2026
Sep '25 Paper accepted to NeurIPS 2025, see you in San Diego!
Sep '25 Joined Amazon as Applied Scientist Intern
Sep '25 Invited reviewer for ICLR 2026
Aug '25 Passed PhD Oral Qualification Exam
Jul '25 Two papers accepted to ACM MM 2025
Jun '25 One paper accepted to USENIX Security 2025
May '25 Joined KLA Corporation as ML Research Intern
May '25 One paper accepted to ICML 2025

Featured Research

MM-Plan workflow
AI Safety · Red-Teaming

Visual Exclusivity Attacks: Automatic Multimodal Red Teaming via Agentic Planning

Under Review · Work done at Amazon

Introduces Visual Exclusivity, a threat model where harm emerges only through visual reasoning. MM-Plan uses agentic planning to achieve 2-5x improvement over baselines on 8 frontier MLLMs including GPT-5 and Claude 4.5 Sonnet. Includes VE-Safety, a 440-instance benchmark across 15 safety categories.

arXiv Code Project Dataset
AReS workflow
MaaS · On-Device

Prime Once, then Reprogram Locally: Efficient Alternative to Black-Box Model Reprogramming

CVPR 2026

A two-stage framework for MaaS VLM adaptation: global priming establishes transferable prompts once, then local reprogramming adapts per instance. Eliminates repeated API calls while matching or exceeding full-access methods across 12 benchmarks.

arXiv Code
DPCore framework
Continual Learning · TTA

DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation

ICML 2025

A lightweight prompt-based method for adapting pre-trained models to continually shifting domains at test time, without source data, labels, or rewards. Dynamic prompt coresets preserve past knowledge while adapting efficiently to new distributions.

arXiv Code

Other Publications

* = equal contribution
Preprint

Continual Test-Time Adaptation: A Comprehensive Survey

S. Maharana*, S. Mishra*, Y. Zhang*, et al.

PDF
Preprint

Agents in the Wild: Safety, Society, and the Illusion of Sociality on Moltbook

Y. Zhang*, et al.

PDF Code Project
Preprint

Adapting in the Dark: Towards Stable and Efficient Black-Box Test-Time Adaptation

Y. Zhang, et al.

PDF Code
TMLR '26

Prompt-based Adaptation in Large-scale Vision Models: A Survey

X. Xiao*, Y. Zhang*, L. Zhao*, et al.

PDF Awesome List
ICASSP '26

CTR-LoRA: Curvature-Aware and Trust-Region Guided Low-Rank Adaptation for LLMs

Z. Wang, M. Mo, X. Xiao, C. Liu, C. Ma, Y. Zhang, et al.

PDF
NeurIPS '25

Doctor Approved: Generating Medically Accurate Skin Disease Images via AI-Expert Feedback

J. Wang*, Y. Zhang*, et al.

PDF Code
ACM MM '25

Visual Instance-aware Prompt Tuning

X. Xiao*, Y. Zhang*, X. Li, T. Wang, X. Wang, Y. Wei, J. Hamm, M. Xu

PDF
ACM MM '25

eSkinHealth: A Multimodal Dataset for Neglected Tropical Skin Diseases

J. Wang, X. Hu, Y. Zhang, et al.

PDF Code
USENIX '25

SoK: Can Synthetic Images Replace Real Data? A Survey of Utility and Privacy

Y. Cheung, Y. Zhang, N. Marrouche, J. Hamm

PDF
WACV '25

OT-VP: Optimal Transport-guided Visual Prompting for Test-Time Adaptation

Y. Zhang, A. Mehra, J. Hamm  Oral

PDF Code
NeurIPS '24

Understanding the Transferability of Representations via Task-Relatedness

A. Mehra, Y. Zhang, J. Hamm

PDF Code

Full list on Google Scholar

Experience

Oak Ridge National Laboratory
Research Scientist Intern
Jan 2026 –Apr 2026

Uncertainty Quantification for Large Foundation Models and Multi-modal LLMs

Amazon
Applied Scientist Intern
Sep 2025 –Dec 2025

Multi-modal LLM RL-based post-training, safety alignment, and automatic red-teaming via agentic planning

KLA Corporation
ML Research Intern
May 2025 –Aug 2025

Vision Foundation Model knowledge distillation for compact semiconductor defect detection

Education

Tulane University
Ph.D. in Computer Science · Advisor: Prof. Jihun Hamm
2022 –Present
Osaka University
B.S. in Mathematics · Award of Excellence · Multiple Merit Scholarships
2018 –2022
Peking University Withdrawn
Withdrew after two years to pursue an independent academic path abroad
2013 –2015

Honors & Awards

2025
Lambda Research Grant
Lambda
2025
NeurIPS Top Reviewer
NeurIPS 2025
2024
Outstanding Teaching Assistant
Tulane University
2019
Award of Excellence
Osaka University
2018–21
Government & Merit Scholarships
Ministry of Education, Japan
2021
Ichikawa International Scholarship
Osaka University

Academic Service

Conference Reviewer

NeurIPS 2025 (Top Reviewer) ICLR 2025–2026 CVPR 2026 ECCV 2026 ICML 2026 ACM MM 2025 WACV 2025

Journal Reviewer

Transactions on Machine Learning Research (TMLR) J. Imaging Informatics in Medicine

Community

NeurIPS 2023 Outreach Volunteer