I am open to collaborating with remote students and visitors (with or without research experience) in the following areas: ^_^
Reliable Foundation Models: Efficient training, tuning, and compression under imperfect data conditions.
Agentic AI: Building robust and generalizable AI agents for real-world tasks.
AI for Science & Bio-Medical: Unlocking large models for data-scarce and high-stakes domains.
/
/
/
/
Biography & Research
My research interests lie in developing reliable and generalizable Artificial Intelligence for real-world applications. While modern AI thrives on massive, pristine datasets, the real world is inherently messy. The ultimate goal of my research is to bridge the gap between idealized lab settings and the chaotic reality of deployment by building models that are robust to data imperfections.
My work centers on the reality that real-world data is often flawed. The central question driving my research is: "How can we design efficient learning algorithms that do not just survive, but thrive, under imperfect conditions?" I focus on addressing four fundamental challenges of real-world data: low quantity, low quality, and skewed distributions, all while maintaining learning efficiency in the era of Large Models.
My research helps realize these goals by making progress in the following four directions, specifically tailored for the era of Large Foundation Models:
Learning from Low-Quality Data
Developing robust frameworks to separate signal from noise when dealing with corrupted or unreliable data. This is crucial for preventing foundation models from memorizing noisy pre-training data or failing during downstream adaptation.
Designing architectures and adaptation strategies that unlock the power of large foundation models in data-scarce, high-stakes domains like medical imaging, achieving high performance even when massive datasets are unavailable.
Ensuring models generalize fairly and effectively across imbalanced and long-tailed data distributions, mitigating the representational biases that large models frequently absorb from real-world, uncurated data.
Improving fundamental model efficiency, knowledge distillation, and architecture design. My goal is to compress the capabilities of large foundation models into efficient forms, making their training and real-world deployment practical and sustainable.
Call for CSC PhD &
Joint-Training (Visiting PhD) Students
Topics: Efficient & Trustworthy Foundation Models, Robust Earth
Observation (EO), Agentic AI & Safety, AI for Science, Reliable/Green Training.
What we offer: Co-supervision with partners (e.g., UNC, ELLIS
Institute Tübingen), access to Exeter HPC (A100), Isambard-AI (5000 H200) &
national resources, supportive publication mentorship.
Funding paths: China Scholarship Council (CSC) full PhD; Exeter–CSC
joint program; 6–24 month joint-training/visiting PhD via CSC or home grants.
How to apply: Email your CV, transcripts. Use the
subject: “CSC PhD / Joint Training – Your Name”.
We welcome emails from prospective students. Feel free to introduce yourself.
News
[Mar. 2026]Research Grant
Got a grant from NVIDIA Academic Grant Program.
[Mar. 2026]Research Grant
Got a grant from Isambard-AI with 10000 GPU hours.
[Feb. 2026]CVPR 2026
TWO paper accepted.
Main Paper:
Confusion-Aware Spectral Regularizer for Long-Tailed Recognition
and
Findings Paper:
SCOPE: Scene-Contextualized Incremental Few-Shot 3D Segmentation.
[Jan. 2026]ICLR 2026
One paper accepted:
Dual-Kernel Adapter: Expanding Spatial Horizons for Data-Constrained Medical Image Analysis.
[Jan. 2026]IEEE Transactions on Image Processing
One papers accepted:
StealthMark: Harmless and Stealthy Ownership Verification for Medical Segmentation via Uncertainty-Guided Backdoors.
[Jan. 2026]CPAL 2026
TWO papers accepted:
Dual-Kernel Adapter & PASS.
[Jan. 2026]ICPR 2026I serve as Area Chair of ICPR 2026..
[Jan. 2026]ICASSP 2026
One paper accepted:
AUDIO DEEPFAKE DETECTION AT THE FIRST GREETING: “HI!”.
[Nov. 2025]CPAL 2026I serve as Area Chair of CPAL 2026..
[Nov. 2025]AAAI 2026
One paper accepted:
TimeCAP: A Channel-Aware Pre-Training Framework for Multivariate Time
Series Forecasting.
[Oct. 2025]ELLIS Society
Joined as an ELLIS Member – grateful to my endorsers,
collaborators, and students.
[Sep. 2025]ACM WSDM 2025
One paper accepted:
SARC: Sentiment-Augmented Deep Role Clustering for Fake News
Detection.
[Sep. 2025]NeurIPS 2025
One paper accepted:
REOBench.
Invited Journal Reviewer: IEEE Transactions on Industrial Informatics,
Wireless Communications and Mobile Computing, ACM Transactions on
Intelligent Systems and Technology
