PeopleofSuperintelligence

Researcher

Lean Wang

Also known as 王乐安

Lean Wang is a language-model researcher whose methods are part of DeepSeek's open models. As first author of "Label Words are Anchors", he showed how label words in in-context examples gather information and steer a model's predictions, and the paper won the Best Long Paper Award at EMNLP 2023. After a summa cum laude B.S. in intelligence science and technology at Peking University, he began a PhD in Xu Sun's LANCO group in 2023 and joined DeepSeek-AI as a research intern in January 2024. His Loss-Free Balancing method for mixture-of-experts routing went into DeepSeek-V3, and his name appears on DeepSeek-V2, DeepSeek-R1, DeepSeek-V3.2 and DeepSeek-V4. He contributes to DeepSeek's model research, most recently as an author of DeepSeek-V4.1-Flash in September 2026.

Badges

Timeline

  1. 2019–2023

    Intelligence science at Peking University

    He studies intelligence science and technology at Peking University from 2019 to 2023 and graduates with a B.S. summa cum laude. Among the department's 2019 entering class, he ranks first.

  2. May 2023–December 2023

    Best long paper at EMNLP 2023

    With Xu Sun's team at Peking University and the WeChat AI team, where he interns, he writes "Label Words are Anchors" as first author and posts it in May 2023. The paper shows that label words in in-context examples collect semantic information in shallow layers and anchor a model's final predictions. In December 2023, it wins the Best Long Paper Award at EMNLP 2023 in Singapore.

  3. 2023–2024

    PhD in Xu Sun's LANCO group

    In 2023, he begins a PhD at Peking University's Institute of Computational Linguistics, in the LANCO group supervised by Xu Sun. His equal-contribution paper "Towards Codable Watermarking for Injecting Multi-bits Information to LLMs" appears at ICLR 2024; its Balance-Marking method lets watermarks in model-generated text carry multi-bit information.

  4. January 2024–May 2024

    Research intern at DeepSeek

    DeepSeek-AI takes him on as a research intern in January 2024. In May 2024, he is among the authors of DeepSeek-V2, a mixture-of-experts language model with 236 billion total parameters, 21 billion of them active for each token.

  5. August 2024–December 2024

    Loss-Free Balancing in DeepSeek-V3

    In August 2024, he is first author of "Auxiliary-Loss-Free Load Balancing Strategy for Mixture-of-Experts", which balances expert load by adding an expert-wise bias to each expert's routing scores. DeepSeek-V3, a 671-billion-parameter model described in a December 2024 report that lists him among its authors, adopts this auxiliary-loss-free strategy.

  6. January 2025–September 2025

    DeepSeek-R1 in Nature

    He is among the authors of DeepSeek-R1, posted in January 2025, which shows that pure reinforcement learning can draw out reasoning abilities in large language models without human-labelled reasoning traces. Nature publishes the work in September 2025, listing him with the DeepSeek-AI Team in Hangzhou.

  7. December 2025–April 2026

    DeepSeek-V3.2 and DeepSeek-V4

    His name appears on DeepSeek-V3.2 in December 2025, a model that introduces DeepSeek Sparse Attention and a scalable reinforcement learning framework. In April 2026, he is among the authors of the DeepSeek-V4 preview, whose Pro model has 1.6 trillion parameters and supports a context of one million tokens.

  8. September 2026

    Coauthoring DeepSeek-V4.1-Flash

    In September 2026, he is among the DeepSeek-AI authors of DeepSeek-V4.1-Flash, a multimodal mixture-of-experts model with 552 billion backbone parameters. Cross-layer cache reuse and FP4 caching bring its global KV cache down to 890 bytes per token.

Career

  1. 2024DeepSeekResearch intern
  2. 2023WeChat AI teamIntern

Education

  1. 2023Peking UniversityB.S.
  2. From 2023Peking UniversityPhD

From the biography.

  1. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes Lean Wang. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.
  2. The median h-index is 48 at universities and 14 at frontier AI labsIncludes DeepSeek. 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. 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.

Colleagues

People with dated positions at the same organization in the same years.

Sources and credits

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

  1. About - Lean Wang’s Homepage leanwang326.github.io. Checked 20 October 2026.
  2. EMNLP 2023 最佳论文放榜！北京大学联合微信AI团队获国内首篇 EMNLP 最佳长论文 cloud.tencent.com. Checked 20 October 2026.
  3. [2305.14160] Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learning arxiv.org. Checked 20 October 2026.
  4. EMNLP 2023 Best Paper公布，最佳长论文奖一作即将在智源社区分享 - 智源社区 hub.baai.ac.cn. Checked 20 October 2026.
  5. [2307.15992] Towards Codable Watermarking for Injecting Multi-bits Information to LLMs arxiv.org. Checked 20 October 2026.
  6. [2405.04434] DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model arxiv.org. Checked 20 October 2026.
  7. [2408.15664] Auxiliary-Loss-Free Load Balancing Strategy for Mixture-of-Experts arxiv.org. Checked 20 October 2026.
  8. [2412.19437] DeepSeek-V3 Technical Report arxiv.org. Checked 20 October 2026.
  9. [2501.12948] DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning arxiv.org. Checked 20 October 2026.
  10. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning | Nature nature.com. Checked 20 October 2026.
  11. [2512.02556] DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models arxiv.org. Checked 20 October 2026.
  12. [2606.19348] DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence arxiv.org. Checked 20 October 2026.
  13. [2609.19969] DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 20 October 2026.