PeopleofSuperintelligence

Researcher

Karl Cobbe

Karl Cobbe is a research scientist at OpenAI whose work spans generalization in reinforcement learning and mathematical reasoning in language models. After studying computer science at Stanford University from 2010 to 2014 and working as a software engineer at Chatous and Art of Problem Solving, they joined OpenAI as a machine learning fellow in 2018. There they released CoinRun in 2018 and the Procgen Benchmark in 2019, two benchmarks for measuring how well reinforcement learning agents generalize, and introduced Phasic Policy Gradient in 2020. In 2021 they introduced GSM8K, a dataset of 8.5K grade school math word problems, and in 2023 they coauthored "Let's Verify Step by Step," which found that process supervision significantly outperforms outcome supervision on the MATH dataset. They were also a foundational contributor to reasoning research for OpenAI o1 and coauthored its system card in December 2024.

Now at
OpenAI
Badges

Timeline

  1. 2010–2017

    Stanford and Art of Problem Solving

    They study computer science at Stanford University as a BS student from 2010 to 2014. Software engineering jobs follow at Chatous from 2014 to 2015 and at Art of Problem Solving from 2016 to 2017.

  2. 2018–2019

    CoinRun at OpenAI

    They join OpenAI in 2018 as a machine learning fellow and become a research scientist there. In December 2018, they release CoinRun, a benchmark for generalization in reinforcement learning that measures an agent's ability to transfer its experience to novel situations. The ICML 2019 paper, written with Oleg Klimov, Chris Hesse, Taehoon Kim and John Schulman, finds that agents overfit to surprisingly large training sets.

  3. 2019–2020

    The Procgen Benchmark

    In December 2019, with Christopher Hesse, Jacob Hilton and John Schulman, they release the Procgen Benchmark, 16 procedurally generated game-like environments that benchmark sample efficiency and generalization in reinforcement learning. The paper appears at ICML 2020, and they give an introduction to the benchmark in the NeurIPS 2020 competition track.

  4. 2020–2021

    Phasic Policy Gradient

    Phasic Policy Gradient, which they introduce in September 2020 with Jacob Hilton, Oleg Klimov and John Schulman, separates policy and value function training into distinct phases. Compared with PPO, the method significantly improves sample efficiency on the Procgen Benchmark, and the paper appears at ICML 2021.

  5. 2021

    GSM8K and WebGPT

    In October 2021, with coauthors including Vineet Kosaraju and John Schulman, they introduce GSM8K, a dataset of 8.5K grade school math word problems, and train verifiers to judge the correctness of model completions. WebGPT follows in December 2021; they coauthor the paper, which fine-tunes GPT-3 to answer long-form questions using a text-based web-browsing environment.

  6. 2023–2024

    Process supervision and PRM800K

    In May 2023, they coauthor "Let's Verify Step by Step" with Hunter Lightman, Vineet Kosaraju, Jan Leike, John Schulman, Ilya Sutskever and others. The study finds that process supervision, which gives feedback on each intermediate reasoning step, significantly outperforms outcome supervision on the MATH dataset, and releases PRM800K, a dataset of 800,000 step-level human feedback labels. The paper appears at ICLR 2024.

  7. 2024

    Foundational contributor to o1

    They are one of the foundational contributors in reasoning research for OpenAI o1. In December 2024, they coauthor the OpenAI o1 System Card, which describes a model series trained with large-scale reinforcement learning to reason using chain of thought.

  8. As of October 2026

    Research scientist at OpenAI

    They have been a research scientist at OpenAI since 2018, focusing on reinforcement learning and AI reasoning. Their work there includes first authorship of GSM8K and contributions to OpenAI's math-reasoning and verifier research.

Career

  1. Since 2018OpenAIMachine learning fellow
  2. 2016–2017Art of Problem SolvingSoftware engineer
  3. 2014–2015ChatousSoftware engineer

Education

  1. 2014Stanford UniversityBS

From the biography.

  1. Among employers with 30+ full-time cases, OpenAI leads primary technical role H-1B pay at $310,000, and Amazon, Google and Microsoft file 57% of all casesIncludes OpenAI. Across 440,669 certified US H-1B filings from 44 employers (FY2020–FY2026), OpenAI leads primary technical role base pay at $310,000 and xAI leads overall base pay at $290,000.
  2. 22 of 133 people shaping AI with a former OpenAI role now work at AnthropicIncludes OpenAI. Our estimates of where OpenAI recruits came from and where 133 people with former OpenAI roles work now, including the companies and labs they founded.
  3. The median h-index is 48 at universities and 14 at frontier AI labsIncludes OpenAI. 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.
  4. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes OpenAI. 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 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 Karl Cobbe yet. Send us a free-licensed photo.

  1. Karl Cobbe | alphaXiv alphaxiv.org. Checked 20 October 2026.
  2. Quantifying Generalization in Reinforcement Learning (OpenAI) web.archive.org. Checked 20 October 2026.
  3. Quantifying Generalization in Reinforcement Learning (PMLR 97) proceedings.mlr.press. Checked 20 October 2026.
  4. Procgen Benchmark | OpenAI web.archive.org. Checked 20 October 2026.
  5. Leveraging Procedural Generation to Benchmark Reinforcement Learning (PMLR 119) proceedings.mlr.press. Checked 20 October 2026.
  6. NeurIPS Introduction to the Procgen Benchmark neurips.cc. Checked 20 October 2026.
  7. [2009.04416] Phasic Policy Gradient arxiv.org. Checked 20 October 2026.
  8. Phasic Policy Gradient (PMLR 139) proceedings.mlr.press. Checked 20 October 2026.
  9. [2110.14168] Training Verifiers to Solve Math Word Problems arxiv.org. Checked 20 October 2026.
  10. Solving math word problems | OpenAI web.archive.org. Checked 20 October 2026.
  11. [2112.09332] WebGPT: Browser-assisted question-answering with human feedback arxiv.org. Checked 20 October 2026.
  12. [2305.20050] Let's Verify Step by Step arxiv.org. Checked 20 October 2026.
  13. Improving mathematical reasoning with process supervision | OpenAI web.archive.org. Checked 20 October 2026.
  14. Cobbe, Karl | ML Anthology mlanthology.org. Checked 20 October 2026.
  15. OpenAI o1 Contributions | OpenAI web.archive.org. Checked 20 October 2026.
  16. [2412.16720] OpenAI o1 System Card arxiv.org. Checked 20 October 2026.
  17. Karl Cobbe - OpenAI research scientist sophon.at. Checked 20 October 2026.