Dingzhi Yu (余 定 之) M.Sc. student, LAMDA Group Supervisor: Professor Lijun Zhang Major: Computer Science and Technology School of Artificial Intelligence State Key Laboratory of Novel Software Technology Nanjing University, Nanjing 210023, China Google Scholar

About Me

I am a final-year graduate student of School of Artificial Intelligence in Nanjing University and a member of LAMDA Group, led by Professor Zhi-Hua Zhou. Prior to that, I received my B.E. degree from School of Data Science, Fudan University in June 2024. Currently, I am a research intern at the University of Illinois Urbana-Champaign, under the valued supervision of Professor Tong Zhang. I also work closely with Professors Difan Zou and Luo Luo.

I work on the interplay between optimization theory and modern machine learning, with the current focus on:

1. designing efficient and scalable optimizers for pretraining and post-training of Large Language Models;


2. explaining the practical efficiency of optimization algorithms from a theoretical perspective.

My featured work ⭐ takes a small step toward these goals. Feel free to reach out if you are also excited about optimization and LLMs.

Preprints

1. StoSignSGD: Unbiased Structural Stochasticity Fixes SignSGD for Training Large Language Models [arXiv, blog] ⭐
   Dingzhi Yu*, Rui Pan*, Yuxing Liu*, Difan Zou, and Tong Zhang (* denotes equal contribution)
   
   
2. Better Convergence Guarantees for Sign-Based Momentum Methods [arXiv]
   Wei Jiang, Dingzhi Yu, Sifan Yang, Wenhao Yang, Zechao Li, and Lijun Zhang
   
   
3. Improved Approximate Regret for Decentralized Online Continuous Submodular Maximization via Reductions [arXiv]
   Yuanyu Wan, Yu Shen, Dingzhi Yu, Bo Xue, and Mingli Song

Publication

1. Sign-Based Optimizers Under Heavy-Tailed Noise: On the effectiveness of Lion and Muon [arXiv, Code] ⭐
   Dingzhi Yu, Hongyi Tao, Yuanyu Wan, and Lijun Zhang
   In Advances in Neural Information Processing Systems 39 (NeurIPS 2026), to appear, 2026.
   
   
2. When and Why SignSGD Outperforms SGD: A Theoretical Study Based on $\ell_1$-norm Lower Bounds [arXiv, Code]
   Hongyi Tao*, Dingzhi Yu*, and Lijun Zhang
   In Advances in Neural Information Processing Systems 39 (NeurIPS 2026), to appear, 2026.
   
   
3. Group Distributionally Robust Optimization with Flexible Sample Queries [arXiv]
   Haomin Bai, Dingzhi Yu, Shuai Li, Haipeng Luo, and Lijun Zhang
   In Advances in Neural Information Processing Systems 39 (NeurIPS 2026), to appear, 2026.
   
   
4. Mirror Descent Under Generalized Smoothness [Link]
   Dingzhi Yu, Wei Jiang, Hongyi Tao, Yuanyu Wan, and Lijun Zhang
   In Proceedings of the 43rd International Conference on Machine Learning (ICML 2026), pages 150427-150481, 2026.
   
   
5. Efficient Algorithms for Empirical Group Distributionally Robust Optimization and Beyond [Link]
   Dingzhi Yu, Yunuo Cai, Wei Jiang, and Lijun Zhang
   In Proceedings of the 41st International Conference on Machine Learning (ICML 2024), pages 57384-57414, 2024.
   
   

Technical Notes

1. Open Problem: Is AdamW Effective Under Heavy-Tailed Noise? [arXiv]
   Dingzhi Yu, Hongyi Tao, Yuanyu Wan, Luo Luo, and Lijun Zhang
   
   
2. Finite-Time Nonsmooth Convergence of Signum and Muon [blog, PDF, bibtex]
   Dingzhi Yu, Yuxing Liu, Rui Pan, Lijun Zhang, Yang You and Difan Zou
   
   

Grant

• LLM Optimizer in Resource Constrained Environments.
  Postgraduate Research & Practice Innovation Program of Jiangsu Province (NO.26CXJH0364)

Awards & Honors

• National Scholarship, 2026
• Xiaomi Outstanding Scholarship (Top-tier Award, ￥20,000), 2025
• Excellent Graduate Student of Nanjing University, 2024
• First-Class Academic Scholarship for Graduate Students (￥10,000), 2024

Academic Service

• Reviewer: JMLR, ICLR, NeurIPS.

Correspondence

• Email: yudz@lamda.nju.edu.cn

• Address: State Key Laboratory of Novel Software Technology, Nanjing University, Xianlin Campus Mailbox 603, 163 Xianlin Avenue, Qixia District, Nanjing 210023, China.