Gan Luo

Gan Luo

Peking University

I’m an incoming PhD student at Stanford [MS&E]. I received my bachelor’s degree in Mathmematics from Peking University in 2026, where I was very fortunate to be advised by Prof. [Kun Yuan]. I have the privilege of working with (time order) Prof. [David Simchi-Levi] at MIT, Prof. [Jing Dong] at Columbia University, Prof. [Wotao Yin] at the [Decision Intelligence Lab, DAMO Academy] and Prof. [Bin Dong] at PKU. Throughout my research journey, I have been fortunate to learn from and collaborate with [Liyuan Liang], [Ruicheng Ao] and [Zihan Qin].

Please feel free to reach out!

📝 Publications and PrePrints

* denotes equal contribution; ($\alpha$-$\beta$) denotes alphabetical order; expand TL;DR for details.

  • Service-Induced Congestion in Memory-Constrained LLM Serving [Arxiv] [SSRN]
    ($\alpha$-$\beta$) Ruicheng Ao*, Jing Dong, Gan Luo*, David Simchi-Levi
    Under review at Operations Research
    Finalist in INFORMS Service Science Section Best Student Paper Award, 2026

  • Achieving Linear Speedup and Optimal Complexity for Decentralized Optimization over Row-stochastic Networks [Arxiv] [Code]
    Liyuan Liang*, Xinyi Chen*, Gan Luo*, Kun Yuan
    ICML 2025 Spotlight (top 2.6%) [Link]
    TL;DR: We determine the optimal convergence rate for decentralized optimization over a row stochastic network and design an algorithm to achieve it.
    • Introduced novel metrics to characterize the influence of row-stochastic mixing matrices and established the first convergence lower bound for decentralized optimization over row-stochastic networks. Developed a new analysis framework proving that [PULL-DIAG algorithm] achieves linear speedup and proposed a multi-gossip protocol that resolves instability issues and attains the lower bound with near-optimal complexity.
    • Conducted all numerical experiments to validate the theoretical results on convergence lower bound, linear speedup, and near-optimal complexity.

  • From Search to Synthesis: Training LLMs as Zero-Shot Workflow Generators [Arxiv]
    Gan Luo*, Zihan Qin*, Bin Dong, Wotao Yin

  • Optimizing LLM Inference: Fluid-Guided Online Scheduling with Memory Constraints [Arxiv] [SSRN] [Code1] [Code2: Simulator] [Demo]
    ($\alpha$-$\beta$) Ruicheng Ao*, Gan Luo*, David Simchi-Levi, Xinshang Wang
    Major Revision at Operations Research
    Preliminary version accepted at NeurIPS 2025 MLxOR Workshop
    Finalist in INFORMS Applied Probability Society Best Student Paper Prize, 2025 [Link] [Award]

  • On the Linear Speedup of the Push-Pull Method for Decentralized Optimization over Digraphs [Arxiv] [Code: Linear Speedup] [Notes]
    Liyuan Liang*, Gan Luo*, Kun Yuan
    Major Revision at Siam Optimization
    TL;DR: We propose a novel multi-step descent analysis framework and first to prove that the [Push-Pull algorithm] achieves linear speedup over arbitrary strongly connected digraphs.
    • This is my first research project, started in December 2023, advised by Prof. Kun Yuan and in collaboration with Liyuan Liang, whom I am fortunate to learn from. At first, we wanted to use the [traditional approach] to analyze the problem, but after several months we found that this would give a non-vanishing noise term in the upper bound. We further found that this was because we first analyzed the single-term noise and then added them together, which motivated us to try a multi-step descent analysis framework, and finally we succeeded at about October 2024. You can see how we clearly analyze the multi-step term in the [notes].
    • For the empirical experiments, we struggled when validating the linear speedup properties of neural networks on the MNIST dataset. The challenge was in correctly measuring the gradient norm: we needed to compute the normalized gradient norm by averaging the gradients from all n nodes (where each node computes its gradient on its own batch) and then normalizing by the square root of the total parameter count, rather than using the average of the individual node gradient norms. See the [implementation]

🎖 Slected Honors and Awards

  • Finalist in INFORMS Applied Probability Society Best Student Paper Prize, 2025. [Link] [Award]
  • Applied Mathematics Elite Program, the program accepted only 15 people this year.
  • Silver Medal, 35th Chinese Chemical Olympiad (National Final). I still find the methodology of total synthesis quite fascinating.

📖 Educations and Experience

  • 2023.04 - 2026.06, Undergraduate Student, School of Mathematical Sciences, Peking University
  • 2025.07 - 2025.09, Intern, DAMO Academy, Hangzhou, China
  • 2022.09 - 2023.04, Undergraduate Student, College of Environmental Sciences and Engineering, Peking University
  • 2019.09 - 2022.06, Senior High School Student, Chongqing Nankai Secondary School

😉 Miscs

  • My nickname is little_wolf, which comes from a Chinese pun on my name. wolf = lang_gou = luo_gan.
  • My favorite color is light blue, rgb(99,142,201).
  • I sometimes write blogs on [Zhihu] and [Personal Blog]. 何もいらない~