|
Xiang Zheng (郑想) Xiang Zheng received her B.Eng. degree in Artificial Intelligence from the South China University of Technology. She is currently working on improving the efficiency and interpretability of diffusion models at Westlake University. She was a visiting student at UC Berkeley, where she worked as an undergraduate research assistant in Prof. Michael Mahoney's group, under the supervision of Prof. Michael Mahoney and Dr. Amir Gholami. Earlier, she interned at the AI for Scientific Simulation and Discovery Lab at Westlake University, mentored by Prof. Tailin Wu. Before that, she conducted research at the Cyber-Med Laboratory, SCUT, under the supervision of Prof. Zhanpeng Jin and Prof. Yang Gao. Email: xiangzhengzx@outlook.com Github: https://xiangzheng2002.github.io/ Tel/Wechat: +86-15355021171 |
|
|
||||||||||||||
|
|
|
Developing discrete diffusion models with improved efficiency and interpretability. Advised by Prof. Tailin Wu. |
||
|
Developed foundation models for scientific computing and SciML agents for automatic code generation. Advised by Prof. Michael Mahoney and Dr. Amir Gholami. |
||
|
|
AI for PDE and physical system simulation and control. Advised by Prof. Tailin Wu. |
||
|
Biometric recognition in human-computer interaction. Advised by Prof. Zhanpeng Jin and Prof. Yang Gao. |
|
|
|
* indicates equal contribution
|
|
Saarth Gaonkar *, Xiang Zheng *, Haocheng Xi, Rishabh Tiwari, Kurt Keutzer, Dmitriy Morozov, Michael Mahoney, Amir Gholami We explore LLMs as SciML agents that generate executable and numerically valid ODE solvers directly from natural-language problem descriptions. We introduce two new datasets, including a diagnostic adversarial benchmark and a large-scale benchmark of 1,000 diverse ODE tasks, and show that modern LLMs, with guided prompting and proper fine-tuning, can reliably select appropriate solvers and produce correct numerical code. Math-AI Workshop, NeurIPS 2025 |
|
Peiyan Hu *, Rui Wang *, Xiang Zheng, Tao Zhang, Haodong Feng, Ruiqi Feng, Long Wei, Yue Wang, Zhi-Ming Ma, Tailin Wu We propose the Wavelet Diffusion Neural Operator (WDNO), a wavelet-domain diffusion framework for PDE simulation and control that enables robust multi-resolution learning under abrupt dynamics. WDNO outperforms strong baselines on Burgers’, compressible Navier-Stokes, 2D incompressible flows, and the ERA5 climate dataset, achieving a 78% reduction in smoke leakage in challenging 2D control tasks. ICLR 2025 |
|
Long Wei *, Haodong Feng *, Yuchen Yang, Ruiqi Feng , Peiyan Hu , Xiang Zheng, Tao Zhang, Dixia Fan, Tailin Wu We propose CL-DiffPhyCon, an efficient closed-loop diffusion-based control framework that generates real-time control signals via asynchronous denoising across heterogeneous physical time steps. CL-DiffPhyCon achieves superior control performance with significantly improved sampling efficiency on 1D Burgers’ equation and 2D incompressible fluid control tasks. ICLR 2025 |
|
* indicates equal contribution
|
|
Amin Totounferoush *, Xiang Zheng *, Arnur Nigmetov , Shashank Subramanian, Haocheng Xi, Kurt Keutzer, Steffen Staab, Dmitriy Morozov, Michael Mahoney, Amir Gholami We study in-context learning (ICL) for solving dynamical systems and propose scientific data augmentation strategies based on fast approximate solvers to improve accuracy and generalization for both forward and inverse problems. Our methods yield up to 6× accuracy gains with ICON-LM and 78% improvement with GPT-4, significantly enhancing both in- and out-of-distribution performance. |
|
Yang Gao, Yuqin Yang, Xiang Zheng, Yaoshi Chen, Yanzi Zhou, Zhanpeng Jin We propose Raise2Auth, a dual-factor, gesture-based mobile authentication system that leverages users’ natural grasping and lifting motions captured by smartphone IMU sensors for robust and seamless identity verification. Experiments with 25 participants achieve over 97.6% authentication accuracy and a false acceptance rate below 0.1% under both imitation and synthetic attacks. |