Huiwen Wu
Associate Researcher · Beihang University, Hangzhou International Innovation Institute
I build trustworthy foundation models through the lens of numerical analysis, working at the intersection of differential privacy, optimization theory, and AI for science.
My research combines PhD training in randomized solvers (UC Irvine, Mathematics) with industrial-scale deployment experience at Ant Group, and applies it to privacy-preserving large language models and scientific foundation models. The through-line is that the tools of classical numerical analysis — preconditioning, multigrid, randomized sketching — turn out to be exactly what modern machine learning needs when it has to be both private and large.
I have published 10+ first- or corresponding-author papers at top-tier venues including AAAI, WWW and IJCAI, hold authorized US patents in privacy-preserving model training, and serve as PI on a Zhejiang Lab talent grant. I am a core contributor to National Key R&D Programs, including AI-powered galaxy simulation for the Chinese Space Station Telescope.
Research interests
- Differential privacyRDP and GDP accounting, perturbation mechanism design, and privacy–utility trade-offs that survive contact with real models.
- Federated optimizationGradient compression, subspace descent, and personalization for training across parties that cannot share data.
- Numerical analysis for MLPreconditioning, multigrid, and randomized solvers brought to bear on foundation-model training.
- AI for scienceSeismic imaging and galaxy simulation, where the physics constrains what a learned model is allowed to do.
Selected work
Multigrid–ViT Fusion for Seismic Foundation Models
Communication-Efficient and Differentially Private Federated Fine-tuning for LLMs
Differentially Private Knowledge Transfer for Recommendation
Read more about these projects →