PeopleofSuperintelligence

Researcher

Song Mei

Also known as 梅松

Song Mei is an associate professor of statistics and electrical engineering and computer sciences at UC Berkeley who studies the theory behind deep learning and generative AI. He earned a mathematics degree at Peking University in 2014, completed a Stanford doctorate in 2020 with Andrea Montanari, and joined Berkeley's Department of Statistics in 2020. His papers examined neural network training, learning in games with incomplete information, and how transformers perform statistical algorithms from examples. In 2025, he received a Sloan Research Fellowship and the Noether Early Career Scholar Award. While on leave at OpenAI in 2025, he contributed to GPT-5, GPT-5.2, and GPT-oss.

Works
1
Citations
0
h-index
0

Publication numbers from OpenAlex, as of 28 October 2026.

Timeline

  1. 2014

    Mathematics at Peking University

    In 2014, he earns a BS in mathematics at Peking University.

  2. 2018

    A mean field theory

    With Andrea Montanari and Phan-Minh Nguyen, he coauthors A Mean Field View of the Landscape of Two-Layers Neural Networks in 2018. They show how a partial differential equation can describe stochastic gradient descent in a limit of large neural networks.

  3. 2020

    Stanford doctorate and Berkeley appointment

    He completes a doctorate in computational and mathematical engineering at Stanford in 2020, advised by Andrea Montanari. In 2020, he joins UC Berkeley's Department of Statistics as an assistant professor.

  4. 2022

    Learning games with incomplete information

    He coauthors Near-Optimal Learning of Extensive-Form Games with Imperfect Information at ICML in 2022 with Yu Bai, Chi Jin, and Tiancheng Yu. The work introduces Balanced Online Mirror Descent and Balanced Counterfactual Regret Minimization to learn strategies in games where players have incomplete information.

  5. 2023

    Transformers as statisticians

    In 2023, he coauthors Transformers as Statisticians with Yu Bai, Fan Chen, Huan Wang, and Caiming Xiong. They show that transformers can perform algorithms such as least squares and ridge regression from examples supplied in context.

  6. 2024

    A survey of diffusion models

    With Minshuo Chen, Jianqing Fan, and Mengdi Wang, he coauthors An Overview of Diffusion Models in 2024. The survey explains applications, guided generation, statistical properties, and sampling capabilities of these generative models.

  7. 2025

    Sloan and Noether awards

    He receives a Sloan Research Fellowship and the American Statistical Association's Noether Early Career Scholar Award in 2025.

  8. 2025

    GPT models at OpenAI

    In 2025, while on leave at OpenAI, he is a core contributor to GPT-5 and GPT-5.2 and also contributes to GPT-oss.

  9. May 2026

    Internalizing chain of thought

    In May 2026, he coauthors Transformers Provably Learn to Internalize Chain-of-Thought with Yixiao Huang, Hanlin Zhu, Zixuan Wang, Jiantao Jiao, Stuart Russell, and Somayeh Sojoudi. The work studies Log-ICoT, a training curriculum for parity tasks, and shows how transformers can absorb reasoning steps into their hidden states.

  10. As of October 2026

    Associate professor at Berkeley

    As of October 2026, he is an associate professor in UC Berkeley's Department of Statistics and Department of Electrical Engineering and Computer Sciences. He studies language models, diffusion models, deep learning theory, and reinforcement learning theory.

Career

  1. NowUniversity of California, BerkeleyAssociate Professor, Statistics and Electrical Engineering and Computer Sciences
  2. From 2020University of California, BerkeleyAssistant Professor, Department of Statistics

Education

  1. 2020Stanford UniversityPhD in computational and mathematical engineering
  2. 2014Peking UniversityBS in mathematics

From the biography and OpenAlex affiliations (the years of papers that list the organization), so some years are approximate.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Song Mei yet. Send us a free-licensed photo.

  1. Song Mei | EECS at UC Berkeley www2.eecs.berkeley.edu. Checked 28 October 2026.
  2. [1804.06561] A Mean Field View of the Landscape of Two-Layers Neural Networks arxiv.org. Checked 28 October 2026.
  3. Mei Wins Okawa Foundation Grant | Department of Statistics statistics.berkeley.edu. Checked 28 October 2026.
  4. Song Mei (UC Berkeley) bliss.eecs.berkeley.edu. Checked 28 October 2026.
  5. Near-Optimal Learning of Extensive-Form Games with Imperfect Information proceedings.mlr.press. Checked 28 October 2026.
  6. [2306.04637] Transformers as Statisticians: Provable In-Context Learning with In-Context Algorithm Selection arxiv.org. Checked 28 October 2026.
  7. [2404.07771] An Overview of Diffusion Models: Applications, Guided Generation, Statistical Rates and Optimization arxiv.org. Checked 28 October 2026.
  8. Two early-career CDSS faculty named 2025 Sloan Fellows | CDSS at UC Berkeley cdss.berkeley.edu. Checked 28 October 2026.
  9. Gottfried E. Noether Awards amstat.org. Checked 28 October 2026.
  10. Song Mei stat.berkeley.edu. Checked 28 October 2026.
  11. [2605.28600] Transformers Provably Learn to Internalize Chain-of-Thought arxiv.org. Checked 28 October 2026.
  12. Song Mei | Department of Statistics statistics.berkeley.edu. Checked 28 October 2026.