PeopleofSuperintelligence

Researcher

Andrew Zhao

Also known as 赵启晨

Andrew Zhao is an AI researcher whose work links reinforcement learning with language-model agents. He studied at the University of British Columbia and the University of Southern California, interned at Peking University, and completed a PhD at Tsinghua. His papers examined surprise-based exploration, agents that learn from experience, self-reference, self-play and language-model reasoning; one earned a NeurIPS Best Paper Runner-Up award in 2025. Since 2026, he is a researcher at Mistral AI and coauthored Voxtral TTS.

Badges

Timeline

  1. 2012–2021

    Undergraduate and graduate study

    He studies as an undergraduate at the University of British Columbia from 2012 to 2017, pursues an MS at the University of Southern California from 2019 to 2020, and interns at Peking University from 2020 to 2021.

  2. 2021–2026

    PhD at Tsinghua

    He pursues a PhD at Tsinghua University from 2021 to 2026 and completes it in 2026. He is a PhD student with Tsinghua's LEAP Lab.

  3. 2022

    A mixture of surprises

    At NeurIPS 2022, he coauthors A Mixture of Surprises for Unsupervised Reinforcement Learning, which trains one policy to maximize surprise and another to minimize it for reward-free exploration.

  4. 2023

    Learning from self-reference

    In 2023, he coauthors Augmenting Unsupervised Reinforcement Learning with Self-Reference, which uses historical information to improve agent performance during pretraining and fine-tuning.

  5. 2024

    ExpeL and language agents

    In 2024, he researches at Microsoft and the Beijing Institute for General Artificial Intelligence (BIGAI), and coauthors ExpeL with Tsinghua researchers. The agent gathers experiences and extracts natural-language knowledge from training tasks for later recall at inference.

  6. May 2025

    Absolute Zero self-play

    In May 2025, he coauthors Absolute Zero, a self-play method where one model proposes verifiable tasks and improves its reasoning by solving them. Its Absolute Zero Reasoner uses a code executor to validate generated code tasks and verify answers.

  7. 2025

    A NeurIPS paper runner-up

    In 2025, he coauthors a study of RLVR and language-model reasoning. It finds RLVR improves sampling efficiency and that reasoning paths generated by RLVR models already appear in base models' sampling distributions. Base models also reach higher pass@k scores at large k, and NeurIPS names the paper a Best Paper Runner-Up.

  8. Since 2026

    Researcher at Mistral AI

    Since 2026, he is a researcher at Mistral AI. In March 2026, he coauthors Voxtral TTS, a multilingual text-to-speech model.

Career

  1. Since 2026Mistral AIResearcher
  2. 2024Microsoft
  3. 2024Beijing Institute for General Artificial Intelligence
  4. 2020–2021Peking UniversityIntern

Education

  1. 2026Tsinghua UniversityPhD
  2. 2020University of Southern CaliforniaMS
  3. 2017University of British Columbia

From the biography.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Andrew Zhao. 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 Andrew Zhao. 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. Language models reach 33% of 2020s entrants, and company shares run from 8% in pattern recognition to 91% in AI infrastructureIncludes Andrew Zhao. We estimate how research topics in AI changed across seven entry decades (6,853 people from the 1960s to the 2020s) and how 60 major topics split between companies (8% to 91%) and universities (10% to 93%).
  4. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes Mistral AI. 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.

Colleagues

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

Sources and credits

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

  1. Andrew Zhao (Mistral AI): reinforcement learning — Lacuna lacuna.tiptreesystems.com. Checked 21 October 2026.
  2. Gao Huang Homepage gaohuang-net.github.io. Checked 21 October 2026.
  3. Members | LEAP Lab @ Tsinghua University leaplab.ai. Checked 21 October 2026.
  4. ExpeL: LLM Agents Are Experiential Learners arxiv.org. Checked 21 October 2026.
  5. A Mixture Of Surprises for Unsupervised Reinforcement Learning papers.neurips.cc. Checked 21 October 2026.
  6. Augmenting Unsupervised Reinforcement Learning with Self-Reference arxiv.org. Checked 21 October 2026.
  7. ExpeL: LLM Agents Are Experiential Learners andrewzh112.github.io. Checked 21 October 2026.
  8. Absolute Zero: Reinforced Self-play Reasoning with Zero Data arxiv.org. Checked 21 October 2026.
  9. Absolute Zero Reasoner andrewzh112.github.io. Checked 21 October 2026.
  10. Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? papers.nips.cc. Checked 21 October 2026.
  11. Voxtral TTS arxiv.org. Checked 21 October 2026.