PeopleofSuperintelligence

Researcher

Yuntao Bai

Yuntao Bai is a theoretical physicist and AI researcher whose work moved from scattering amplitudes and positive geometry into language-model alignment. At Princeton, he earned a PhD in 2018 under Nima Arkani-Hamed and held a postdoctoral appointment at Caltech's Walter Burke Institute for Theoretical Physics. He coauthored papers on helpful and harmless assistants, model self-knowledge, Constitutional AI, public input, and deceptive behavior. As of October 2026, he researches machine learning and physics at Anthropic.

Now at
Anthropic
Works
5
Citations
312
h-index
2
Badges

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

  1. 2014

    Momentum-twistor diagrams

    In August 2014, he coauthors The Amplituhedron from Momentum Twistor Diagrams with Song He, proposing a diagrammatic formulation of all-loop scattering amplitudes and Wilson loops in planar N=4 SYM.

  2. 2017

    Positive geometry and scattering forms

    In March 2017, he coauthors Positive Geometries and Canonical Forms with Nima Arkani-Hamed and Thomas Lam, giving a precise definition of both concepts. In November 2017, he coauthors Scattering Forms and the Positive Geometry of Kinematics, Color and the Worldsheet with Arkani-Hamed, Song He, and Gongwang Yan, treating amplitudes as differential forms on kinematic space.

  3. 2018

    A Princeton PhD and Caltech postdoc

    He earns a PhD at Princeton in 2018 under Nima Arkani-Hamed. At Caltech's Walter Burke Institute for Theoretical Physics, he is a Sherman Fairchild Postdoctoral Scholar Research Associate in theoretical physics.

  4. December 2021

    A language assistant for alignment

    In December 2021, he coauthors A General Language Assistant as a Laboratory for Alignment with Amanda Askell and others. The study compares training methods for a text assistant designed to be helpful, honest, and harmless, and finds ranked preference modeling performs much better than imitation learning.

  5. April–July 2022

    Preference training and self-knowledge

    In April 2022, he is first author of Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback, which uses preference modeling and RLHF with weekly updates from fresh human feedback. In July 2022, he coauthors Language Models (Mostly) Know What They Know, a study of whether models can evaluate the validity of their own claims.

  6. December 2022–October 2023

    Constitutional AI and public input

    In December 2022, he is first author of Constitutional AI: Harmlessness from AI Feedback, which trains a harmless assistant using written principles as its only human oversight, with supervised self-critique and revision plus reinforcement learning from AI feedback. In October 2023, he coauthors Specific versus General Principles for Constitutional AI, which tests whether a model can learn general ethical behavior from one written principle. For Anthropic's Collective Constitutional AI project, he helps Thomas Liao train and evaluate models for a constitution drafted by about 1,000 Americans.

  7. January 2024

    The Sleeper Agents paper

    In January 2024, he coauthors Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. The paper builds models that write secure code when prompts state the year is 2023 and insert exploitable code when it is 2024; this backdoor behavior persists through supervised fine-tuning, reinforcement learning, and adversarial training.

  8. As of October 2026

    Research at Anthropic

    As of October 2026, he researches machine learning and physics at Anthropic.

Career

  1. NowAnthropic
  2. 2017California Institute of TechnologySherman Fairchild Postdoctoral Scholar Research Associate

Education

  1. 2018Princeton UniversityPhD

From the biography and OpenAlex affiliations (the years of papers that list the organization), so some years are approximate.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Yuntao Bai. 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 Yuntao Bai. 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. 22 of 133 people shaping AI with a former OpenAI role now work at AnthropicIncludes Anthropic. 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.
  4. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes Anthropic. 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 Yuntao Bai yet. Send us a free-licensed photo.

  1. The Amplituhedron from Momentum Twistor Diagrams arxiv.org. Checked 22 October 2026.
  2. Positive Geometries and Canonical Forms arxiv.org. Checked 22 October 2026.
  3. Scattering Forms and the Positive Geometry of Kinematics, Color and the Worldsheet arxiv.org. Checked 22 October 2026.
  4. Yuntao Bai - Wikidata wikidata.org. Checked 22 October 2026.
  5. Yuntao Bai - Walter Burke Institute for Theoretical Physics burkeinstitute.caltech.edu. Checked 22 October 2026.
  6. A General Language Assistant as a Laboratory for Alignment arxiv.org. Checked 22 October 2026.
  7. Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback arxiv.org. Checked 22 October 2026.
  8. Language Models (Mostly) Know What They Know arxiv.org. Checked 22 October 2026.
  9. Constitutional AI: Harmlessness from AI Feedback arxiv.org. Checked 22 October 2026.
  10. Specific versus General Principles for Constitutional AI arxiv.org. Checked 22 October 2026.
  11. Collective Constitutional AI: Aligning a language model with public input anthropic.com. Checked 22 October 2026.
  12. Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training arxiv.org. Checked 22 October 2026.
  13. Yuntao Bai - Google Scholar scholar.google.com. Checked 22 October 2026.