Optimization for AI
Adaptive and matrix-aware optimization methods for efficient, stable large-scale model training.
Research Scientist · Z.ai
Minxin Zhang
章旻昕
I develop optimization algorithms that make large-scale learning more efficient, reliable, and mathematically grounded.
Since September 2026
Research ScientistZ.ai · Beijing, ChinaAbout
I am a Research Scientist at Z.ai in Beijing, where I work on optimization for large-scale AI and language model training.
Previously, I was a Hedrick Assistant Adjunct Professor in the Department of Mathematics at UCLA, where I was fortunate to work with Prof. Hayden Schaeffer and Prof. Deanna Needell.
I earned my Ph.D. in Mathematics at the University of California, San Diego, under the supervision of Prof. Philip E. Gill. I previously received a Master of Mathematics from the University of Waterloo in Canada, under the supervision of Prof. Kirsten Morris, and a B.S. in Mathematics from Wuhan University in China. I also interned at Alibaba's Decision Intelligence Lab (Sunnyvale, CA) in summer 2022.
Research
Research interests
My work combines optimization theory, scalable algorithm design, and computation. Currently, I am especially interested in optimization algorithms for AI and large language model training.
Mathematical optimization
Constrained, nonconvex, stochastic, and global optimization algorithms with rigorous convergence analysis.
Scientific computing
Randomized numerical linear algebra, tensor computation, and optimization for scientific problems.
Selected work
Selected publications
Learned preconditioning for a primal-dual interior-point method
Under review · arXiv:2609.35665
Adam improves Muon: Adaptive moment estimation with orthogonalized momentum
Under review · arXiv:2602.17080
AdaGrad meets Muon: Adaptive stepsizes for orthogonal updates
NeurIPS OPT Workshop · arXiv:2509.02981
Inexact proximal point algorithms for zeroth-order global optimization
Journal of Scientific Computing (accepted) · arXiv:2412.11485
Block matrix and tensor randomized Kaczmarz methods for linear feasibility problems
La Matematica · 4, 557–585
A projected-search interior-point method for nonlinearly constrained optimization
Computational Optimization and Applications · 88(1), 37–70
A class of projected-search methods for bound-constrained optimization
Optimization Methods and Software · 39(3), 459–488
Sensor choice for minimum error variance estimation
IEEE Transactions on Automatic Control · 63(2), 315–330
Teaching & mentoring
Course Instructor & TA
Instructor at UCLA (2023–2026) for PIC 10A: Introduction to Programming (C++), PIC 16A: Python with Applications I, and PIC 16B: Python with Applications II.
I previously served as a teaching assistant at UC San Diego and UWaterloo for courses in optimization, numerical analysis, and applied mathematics.
RIPS & REU research teams
Research in Industrial Projects for Students (RIPS), IPAM · Summer 2025. Academic mentor. Project: Accelerating Linear Programming on AMD GPUs.
Research Experiences for Undergraduates (REU), UCLA · Summer 2024. Mentor. Project: AI for Health and Criminal Justice.
Service
Journal review
- Numerical Algorithms (Springer)
- Numerische Mathematik
- Journal of Machine Learning Research (JMLR)
- Optimization Letters
- IEEE Transactions on Automatic Control (TAC)
- IEEE Transactions on Control Systems Technology (TCST)
Conference & proceedings review
- Neural Information Processing Systems (NeurIPS)
- Sampling Theory and Applications (SampTA)
- AWM Research Symposium proceedings
- IEEE Conference on Decision and Control (CDC)