PeopleofSuperintelligence

Researcher

Kaifeng Lyu

Also known as 吕凯风

Kaifeng Lyu is a machine learning theorist who studies how neural networks train and how to make language-model safety last, and he is a co-author of a 2025 ICLR Outstanding Paper on shallow safety alignment. He founded the Universal Online Judge in 2014, graduated from Tsinghua's Yao Class in 2019, and earned a computer science PhD at Princeton in 2024 under Sanjeev Arora. His papers cover margin maximization in homogeneous networks in 2019, sharpness reduction under normalization in 2022, and grokking in 2023, and a postdoctoral post at UC Berkeley's Simons Institute preceded his move home to Tsinghua. He is an assistant professor at Tsinghua's Institute for Interdisciplinary Information Sciences, where his research covers machine learning theory, AI safety and alignment, and optimization.

Badges

Timeline

  1. 2014

    Founding Universal Online Judge

    He founds the Universal Online Judge (UOJ) in 2014, a popular online judge system in China. UOJ tests traditional and non-traditional programming problems from the Olympiad in Informatics.

  2. 2019

    Yao Class degree at Tsinghua

    He graduates in 2019 with a B.Eng. in computer science from the Yao Class at Tsinghua University.

  3. 2019

    Margin maximization in neural networks

    In June 2019, he and Jian Li post a paper on gradient descent in homogeneous neural networks with ReLU or LeakyReLU activations. They show that once the training loss falls below a threshold, a smoothed version of the normalized margin increases over time.

  4. 2022

    Sharpness reduction with normalization

    With Zhiyuan Li and Sanjeev Arora, he posts a paper in June 2022 on normalization layers and generalization. Their analysis suggests that normalization together with weight decay encourages gradient descent to reduce the sharpness of the loss surface.

  5. 2023

    Explaining grokking in theory

    In November 2023, he posts a paper with Jikai Jin, Zhiyuan Li, Simon S. Du, Jason D. Lee and Wei Hu on grokking. They prove that training can stay at a kernel predictor for a long time before a sharp transition to min-norm or max-margin predictors changes test accuracy.

  6. 2024

    Princeton PhD and Simons postdoc

    He earns a computer science PhD from Princeton University in 2024, advised by Sanjeev Arora. He works as a postdoctoral researcher at the Simons Institute for the Theory of Computing at UC Berkeley before joining Tsinghua.

  7. 2025

    An ICLR Outstanding Paper

    He is one of eight authors of a paper on the depth of safety alignment in language models, named an Outstanding Paper at ICLR 2025. The paper describes shallow safety alignment, in which alignment adapts a model's output mainly over its first few tokens.

  8. Since 2025

    Assistant professor at Tsinghua

    He joins the Institute for Interdisciplinary Information Sciences at Tsinghua University in 2025 as an assistant professor. His research covers machine learning theory, AI safety and alignment, and optimization, and he teaches the course Large Language Models from Scratch: Theory and Practice in 2025 and 2026.

Career

  1. Since 2025Tsinghua UniversityAssistant Professor
  2. University of California, BerkeleyPostdoctoral Researcher

Education

  1. 2024Princeton UniversityPhD
  2. 2019Tsinghua UniversityB.Eng.

From the biography.

  1. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionIncludes Tsinghua University. We estimate where the people shaping AI studied and where their careers led. Massachusetts Institute of Technology leads with 467 former students, including 291 recorded doctorates.
  2. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Tsinghua University. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  3. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes Tsinghua University. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.
  4. 84% of people shaping AI with recorded undergraduate and doctoral schools changed institutions for their doctorateIncludes Tsinghua University. 84% of people shaping AI with recorded undergraduate and doctoral schools changed institutions for their doctorate. Across 2,116 people with both an undergraduate degree and a doctorate recorded, we trace which undergraduate schools feed which doctoral programs, how often students stay at the same school, and how paths differ by birth-country region.

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 Kaifeng Lyu yet. Send us a free-licensed photo.

  1. Kaifeng Lyu kaifeng.ac. Checked 21 October 2026.
  2. Kaifeng Lyu | Simons Institute for the Theory of Computing simons.berkeley.edu. Checked 21 October 2026.
  3. Gradient Descent Maximizes the Margin of Homogeneous Neural Networks arxiv.org. Checked 21 October 2026.
  4. Understanding the Generalization Benefit of Normalization Layers: Sharpness Reduction arxiv.org. Checked 21 October 2026.
  5. Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking arxiv.org. Checked 21 October 2026.
  6. Kaifeng Lyu (Chinese homepage) kaifeng.ac. Checked 21 October 2026.
  7. Safety Alignment Should Be Made More Than Just a Few Tokens Deep arxiv.org. Checked 21 October 2026.
  8. Announcing the Outstanding Paper Awards at ICLR 2025 blog.iclr.cc. Checked 21 October 2026.