Wenjie Zhao.
Student Researcher @ Google · Fall 2026

Wenjie Zhao

I’m a Ph.D. candidate in Computer Science at the University of Texas at Dallas, advised by Prof. Yunhui Guo. My research makes perception models reliable in the open world — detecting out-of-distribution inputs, adapting at test time, and catching hallucinations in vision-language models.

Get in touch CV Google Scholar GitHub
Portrait of Wenjie Zhao

Research focus

01

The world may go beyond what the system has learned, or the system may produce interpretations that go beyond what the visual evidence supports.

what the model has learned unknown
Beyond what it has learnedInputs unlike anything seen in training. The model should notice, not guess with confidence.OOD detection · test-time adaptation
image “A cup on a table, next to a laptop.” no visual evidence model’s description
Beyond what it seesDescriptions that claim more than the image shows. The model should stay grounded.Hallucination detection in LVLMs

Out-of-distribution detection

Spotting and segmenting the unknown — from pixel-level OOD segmentation to continual and survival-analysis settings.

Reliable multimodal models

Hallucination detection in LVLMs, audio-visual segmentation, and egocentric gaze estimation.

Robust & efficient perception

Open-set test-time adaptation, transferability estimation, and efficient segmentation on edge devices.

Recent news

02

  • IndustryJoined Google as a Student Researcher in San Jose, working on efficient moving-object segmentation for edge devices.

  • AwardBest Paper Award at a CVPR 2026 Workshop for our study of future-privileged supervision in egocentric gaze estimation.

  • PaperOpen-set test-time adaptation paper accepted to CVPR 2026 Findings; CURE-OOD accepted to TMLR.

  • IndustryInterned at Epsilon, building a multi-agent LLM system for scalable, replayable data generation.

  • PaperAudio-visual segmentation analysis accepted to AAAI 2026; H2ST accepted to CVPR 2025.

  • TeachingInstructor of Record for ECS 1100: Introduction to Engineering and Computer Science at UT Dallas.

  • IndustryResearch intern at Bosch Research, Pittsburgh.

Experience

03
Sep – Dec 2026
● Current
San Jose, CA

Google

Student Researcher

Research on efficient moving-object segmentation for resource-constrained edge devices.

May – Aug 2026
Irving, TX

Epsilon

Intern

Designed and implemented a multi-agent LLM system for scalable, low-cost, and replayable data generation across multiple industries.

Aug – Dec 2025
Dallas, TX

University of Texas at Dallas

Instructor of Record

Taught ECS 1100: Introduction to Engineering and Computer Science.

May – Aug 2025
Pittsburgh, PA

Bosch Research

Research Intern

Developed CLIP-based models using voltage and program traces for side-channel program inference, and investigated trace-alignment methods for robustness to measurement errors.

Publications

04
Current vs. historical risk for hallucinated and non-hallucinated objects
Under review

Beyond the Current Prediction: Tracking Decoding Trajectories for Object Hallucination Detection in LVLMs

Wenjie Zhao, Jia Li, Ruiyu Mao, Yunhui Guo

Open-set test-time adaptation overview
CVPR 2026 Findings

Mitigating the ID–OOD Tradeoff in Open-Set Test-Time Adaptation

Wenjie Zhao, Jia Li, Xin Dong, Yapeng Tian, Yu Xiang, Yunhui Guo

[paper]
Distribution shift in CT scans across scanners and institutions
TMLR 2026

CURE-OOD: Benchmarking Out-of-Distribution Detection for Survival Prediction

Wenjie Zhao, Jia Li, Mingrui Liu, Jing Wang, Yunhui Guo

[paper]
Controlled study of future-privileged supervision for gaze estimation
CVPR Workshop 2026★ Best Paper

How Much Future Helps? A Controlled Study of Future-Privileged Supervision for Causal Egocentric Gaze Estimation

Jia Li, Wenjie Zhao, Fnu Atisri, Sanskriti Aripineni, Shijian Deng, Jon E. Froehlich, Yuhang Zhao, Yapeng Tian

[paper]
Audio-visual segmentation bias
AAAI 2026

Do Audio-Visual Segmentation Models Truly Segment Sounding Objects?

Jia Li, Wenjie Zhao, Ziru Huang, Yunhui Guo, Yapeng Tian

[paper]
ARGaze autoregressive gaze estimation pipeline
arXiv

ARGaze: Autoregressive Transformers for Online Egocentric Gaze Estimation

Jia Li, Wenjie Zhao, Shijian Deng, Bolin Lai, Yuheng Wu, Ruijia Chen, Jon E. Froehlich, Yuhang Zhao, Yapeng Tian

[paper]
Ad hoc vs. task-aware transferability metric selection
arXiv

MetaRank: Task-Aware Metric Selection for Model Transferability Estimation

Yuhang Liu, Wenjie Zhao, Xin Wang, Yunhui Guo

[paper]
Timeline of out-of-distribution segmentation methods
arXiv 2025

From Pixel to Mask: A Survey of Out-of-Distribution Segmentation

Wenjie Zhao, Jia Li, Yunhui Guo

[paper]
H2ST overview
CVPR 2025

H2ST: Hierarchical Two-Sample Tests for Continual Out-of-Distribution Detection

Yuhang Liu, Wenjie Zhao, Yunhui Guo

[paper]
AV-Mamba architecture
CVPR Workshop 2024

AV-Mamba: Cross-Modality Selective State Space Models for Audio-Visual Question Answering

Ziru Huang, Jia Li, Wenjie Zhao, Yunhui Guo, Yapeng Tian

[paper]
S2M OOD segmentation
CVPR 2024

Segment Every Out-of-Distribution Object (S2M)

Wenjie Zhao, Jia Li, Xin Dong, Yu Xiang, Yunhui Guo

[paper] [code]

Education & honors

05

Education

University of Texas at Dallas

Ph.D. Candidate in Computer Science

Xi’an Jiaotong University

Graduate study in Software Engineering

Sichuan University

Honors B.E. in Automation, Wu Yuzhang Honors College

Honors & support

Best Paper AwardCVPR Workshop 2026
Research compute grantsGoogle Cloud Research Grant · Lambda Research Grant
ReviewerNeurIPS · ICML · ICCV · ACL · MICCAI · MMFM

Toolbox

PythonC/C++PyTorchLaTeXGoogle CloudLambda CloudVision-Language ModelsState Space Models (Mamba)Multi-agent LLMsEdge Computing
Off the clock

When I’m not training models, I’m flying them.

Fixed-wing RC planes, FPV drones, cycling, swimming, the gym, table tennis, hiking — and lately, tennis.

See my life page →