PeopleofSuperintelligence

Researcher

Junxiao Song

Junxiao Song is a principal researcher at DeepSeek AI who proposed GRPO, the reinforcement learning algorithm used to train nearly all DeepSeek models, including DeepSeek-R1. He studied automation at Zhejiang University and completed a PhD in electronic and computer engineering at the Hong Kong University of Science and Technology under Daniel Palomar in 2015, working on optimization methods for signal processing. In industry, he worked on game AI at NetEase Games and then at inspir.ai, where his team built StarCraft Commander, an agent that defeated GrandMaster players in StarCraft II. GRPO appeared in the DeepSeekMath paper in February 2024; he also co-developed DeepSeek-V3 and was a core contributor to DeepSeek-R1, which Nature published in September 2025. In 2026, he coauthored the DeepSeek-V4 and DeepSeek-V4.1-Flash reports.

Now at
DeepSeek
Badges

Timeline

  1. By 2011

    Automation at Zhejiang University

    He studies automation at Zhejiang University and earns his bachelor's degree there in 2011. He also receives the Distinguished Zhu Kezhen Scholarship, the university's highest honour for students.

  2. 2011–2015

    A PhD with Daniel Palomar

    He completes a PhD in electronic and computer engineering at the Hong Kong University of Science and Technology under Daniel Palomar, defending his thesis in July 2015. His research covers optimization methods for signal processing, including two 2014 papers with Prabhu Babu and Palomar: one on the sparse generalized eigenvalue problem, and one on majorization-minimization algorithms that design unimodular sequences with low autocorrelation sidelobes for radar and CDMA.

  3. By 2021

    StarCraft Commander at inspir.ai

    He works at NetEase Games as a senior data mining researcher, responsible for a research team on personalized recommendation and deep reinforcement learning for game AI. He then joins inspir.ai as a research scientist, where his team builds StarCraft Commander, an agent described in an ICML 2021 paper that defeats GrandMaster players with an order of magnitude less computation.

  4. February 2024

    GRPO in DeepSeekMath

    At DeepSeek, he coauthors DeepSeekMath in February 2024, a 7B model that scores 51.7% on the competition-level MATH benchmark. The paper introduces Group Relative Policy Optimization (GRPO), a variant of PPO that he proposes; DeepSeek goes on to apply GRPO to train nearly all its models, including DeepSeek-R1.

  5. 2024

    DeepSeek-V2 and Prover-V1.5

    In May 2024, he is among the authors of DeepSeek-V2, a Mixture-of-Experts language model with 236B total parameters, 21B of them active per token. In August 2024, he coauthors DeepSeek-Prover-V1.5, an open-source model for theorem proving in Lean 4 refined with reinforcement learning from proof assistant feedback.

  6. December 2024–January 2025

    DeepSeek-V3 and DeepSeek-R1

    He co-develops DeepSeek-V3, a 671B-parameter Mixture-of-Experts model described in a December 2024 report that needs 2.788M H800 GPU hours for its full training. In January 2025, he is a core contributor to DeepSeek-R1, which builds reasoning abilities in language models through pure reinforcement learning.

  7. September 2025

    DeepSeek-R1 in Nature

    In September 2025, Nature publishes the DeepSeek-R1 paper, with him among the authors from the DeepSeek-AI Team in Hangzhou. The paper shows that pure reinforcement learning can bring out reasoning in large language models without human-labelled reasoning trajectories.

  8. As of October 2026

    Principal researcher at DeepSeek

    He is a principal researcher at DeepSeek AI. In April 2026, he coauthors DeepSeek-V4, whose V4-Pro model has 1.6T parameters and a one-million-token context. In September 2026, he coauthors DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with 552B backbone parameters.

Career

  1. NowDeepSeekPrincipal Researcher
  2. NetEase GamesSenior Data Mining Researcher
  3. inspir.aiResearch Scientist

Education

  1. 2015Hong Kong University of Science and TechnologyPhD
  2. 2011Zhejiang UniversityBachelor's degree in automation

From the biography.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Junxiao Song. 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.
  2. 16% of AI career moves in the 2020s went from a university to a companyIncludes Junxiao Song. We estimate that 16% of AI career moves in the 2020s went from a university to a company; 29% of moves starting at a university went to a company, and 76% ended the university role.
  3. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes DeepSeek. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  4. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staffIncludes DeepSeek. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staff. We test the estimate against 3,913 named model contributors and dated all-staff records across Google DeepMind, OpenAI and Anthropic, plus six additional frontier-grade labs.

Sources and credits

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

  1. [SIST Seminar] Deep Reinforcement Learning for Game AI: A Case Study in StarCraft II shanghaitech.edu.cn. Checked 20 October 2026.
  2. DeepSeek’s rise shows why China’s top AI talent is skipping Silicon Valley restofworld.org. Checked 20 October 2026.
  3. Junxiao Song | Daniel P. Palomar danielppalomar.com. Checked 20 October 2026.
  4. Sparse Generalized Eigenvalue Problem via Smooth Optimization arxiv.org. Checked 20 October 2026.
  5. Optimization Methods for Designing Sequences with Low Autocorrelation Sidelobes arxiv.org. Checked 20 October 2026.
  6. SCC: an efficient deep reinforcement learning agent mastering the game of StarCraft II arxiv.org. Checked 20 October 2026.
  7. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models arxiv.org. Checked 20 October 2026.
  8. My former PhD student Junxiao Song pushing the limits with DeepSeek danielppalomar.com. Checked 20 October 2026.
  9. DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model arxiv.org. Checked 20 October 2026.
  10. DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search arxiv.org. Checked 20 October 2026.
  11. DeepSeek-V3 Technical Report arxiv.org. Checked 20 October 2026.
  12. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning arxiv.org. Checked 20 October 2026.
  13. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning nature.com. Checked 20 October 2026.
  14. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence arxiv.org. Checked 20 October 2026.
  15. DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 20 October 2026.