Zehai Niu

(he/him)

Assistant Professor · Shenzhen University

Professional Summary

Zehai Niu is an Assistant Professor at Shenzhen University. His research focuses on edge-device inference acceleration for Vision-Language Models (VLM) and VLM interpretability — building VLMs that run fast on consumer devices, and that we actually understand.

He received his PhD from the University of Chinese Academy of Sciences (UCAS) in 2024. From July 2024 to May 2026, he was an AI Researcher at HONOR Device, where he led innovation for the gaming business — including M-DLSS 3.0, the industry’s first NPU-GPU heterogeneous frame interpolation system shipped to consumer phones — and the embodied intelligence stack of RobotPhone. During this period, he worked closely with Google’s Gemma team, Qualcomm, and Alibaba’s Qwen (Tongyi Qianwen) Lab.

He has published 20+ papers at ICCV, ECCV, ACM MM, IEEE TCSVT, IEEE TMM and other top venues, holds 10+ patents, and serves as a reviewer for ICLR, ICCV, ECCV, and ACM MM.

Download CV

Education

PhD in Computer Applied Technology

2019-09-01
2024-07-01

University of Chinese Academy of Sciences (UCAS)

Visiting Researcher

2024-05-01
2024-07-01

National University of Singapore (NUS)

Visiting Researcher

2023-09-01
2024-03-31

Tsinghua University

Interests

VLM Edge-device Inference Acceleration VLM Interpretability Vision-Language Models Embodied Intelligence Efficient Multimodal AI

🔬 Research

My lab works on two tightly coupled problems:

① VLM Edge-device Inference Acceleration. Vision-Language Models are the next wave of consumer AI, but most of them are too heavy to run on phones, AR glasses, or robots. We design model architectures, quantization schemes, KV-cache strategies, and NPU-GPU heterogeneous runtimes that make VLMs fast and energy-efficient on-device — building on hard-won experience shipping M-DLSS 3.0 to millions of HONOR users.

② VLM Interpretability. As VLMs make more decisions inside consumer devices and embodied systems, knowing why they do what they do becomes a safety, trust, and product requirement. We study attention attribution, neuron-level concept localization, and failure-mode discovery for multimodal models.

Together, these two directions answer a question we think will define the next decade of edge AI: how do we make small, fast, trustworthy VLMs that real users want to live with?

🎓 Join the Lab

The lab is open year-round to interns, and we have master’s student slots through Shenzhen University.

I do not believe in students working for an advisor. I believe in students and advisors growing together. What I commit to every person who joins this group:

A real research problem with real stakes. Not toy benchmarks. You will work on things that matter for shipped products and for the frontier of VLM research.
First-author papers + industry exposure. Our work targets ICCV/ECCV/ACM MM/NeurIPS, and through our collaborations you will also see how research turns into products inside Google, Qualcomm, Alibaba, and HONOR.
Career investment. Whether you want to go to a top PhD program, a frontier industry lab, or start your own company — I will spend real time helping you get there. Your future is the project.
Compute, mentorship, and respect. Good GPUs. Weekly 1:1s. No silent treatment, no playing politics.

If you are an undergraduate, master’s, or industry engineer who wants to work on edge VLMs or interpretability, please email me with your CV and a short note on what you want to build. I read every message.

我会竭力培养学生想要的未来,与学生的未来一起进步。

🤝 Industry Collaborations

During my time at HONOR Device (2024 – 2026), I worked in close collaboration with the AI teams at:

Google — Gemma open-model team
Qualcomm — on-device inference and NPU acceleration
Alibaba — Qwen (Tongyi Qianwen) Lab

And shipped M-DLSS 3.0, the industry’s first NPU-GPU heterogeneous frame interpolation system, to HONOR gaming phones — proving that frontier research can move from paper to consumer device in months, not years.

I am open to research collaborations, advisory engagements, and joint-lab arrangements with companies and investors building in edge AI, on-device VLMs, embodied intelligence, and AI agents. Please reach out via email.

Featured Publications

Motion Generation