PeopleofSuperintelligence

Researcher

Collin Burns

Collin Burns is an AI researcher whose work spans language-model evaluation, alignment, and internal model behavior. He coauthored MMLU, ETHICS, and MATH, then studied latent knowledge in language models and whether weak supervision can elicit stronger capabilities. His later research introduced unsupervised skill elicitation and examined unintended behavior from post-training data. He studied machine learning at UC Berkeley with Jacob Steinhardt and Dan Klein and worked at OpenAI and Anthropic. In April 2026, he was selected to lead CAISI and left the role within a week; as of October 2026, he is an AI researcher.

Works
16
Citations
1,678
h-index
11

Publication numbers from OpenAlex, as of 19 October 2026.

Timeline

  1. 2015

    A 5.25-second cube record

    In 2015, he breaks the official world record for a single 3x3 solve at 5.25 seconds.

  2. 2020–2021

    MMLU, ETHICS, and MATH

    In 2020, he coauthors MMLU, a test that covers 57 tasks, and the ETHICS benchmark for moral judgments. In 2021, he coauthors the MATH dataset paper on mathematical problem solving.

  3. 2022–2023

    A Berkeley machine-learning PhD

    In 2023, he is a second-year machine-learning PhD student at UC Berkeley, working with Jacob Steinhardt and Dan Klein. His 2022 paper with Haotian Ye, Dan Klein, and Jacob Steinhardt proposes an unsupervised method that finds latent knowledge in a language model's internal activations.

  4. 2024

    Weak-to-strong generalization

    His 2024 ICML paper tests whether weak models can supervise stronger models. It finds that strong models trained on weak-model labels outperform their weak supervisors on natural-language processing, chess, and reward modeling tasks.

  5. 2025–2026

    Unsupervised elicitation and post-training

    He coauthors Unsupervised Elicitation of Language Models in 2025; the paper uses models' own generated labels to elicit skills without external supervision. In 2026, he coauthors Chunky Post-Training, which studies spurious correlations arising from post-training data.

  6. April 2026

    Tapped to lead CAISI

    In April 2026, he is tapped to lead the Center for AI Standards and Innovation after work at OpenAI and Anthropic. The White House pushes him out four days into the job, and Chris Fall takes over.

  7. As of October 2026

    AI researcher

    As of October 2026, he is an AI researcher after work at OpenAI and Anthropic.

Career

  1. 2026US AI Safety Institute / CAISILead

Education

  1. From 2022University of California, BerkeleyPhD

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

Most-cited works

  1. Measuring Massive Multitask Language Understanding2021 · International Conference on Learning Representations · 435 citationsWith Dawn Song
  2. Measuring Mathematical Problem Solving With the MATH Dataset2021 · 306 citationsWith Dawn Song
  3. Aligning AI With Shared Human Values2020 · 166 citationsWith Dawn Song
  4. GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow2021 · 142 citationsWith Dawn Song
  5. CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review2021 · 98 citations

From OpenAlex, as of 19 October 2026.

  1. Through 2025, first appointments to public AI advisory and safety bodies peaked in 2023Includes Collin Burns. We find 27 first recorded public AI appointments in 2023, the peak among complete years. Our estimates cover 63 people across 10 bodies.
  2. 13 of 79 people shaping AI with a former OpenAI role now work at AnthropicIncludes Anthropic. Our estimates of where OpenAI recruits came from and where people with former OpenAI roles work now.
  3. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionIncludes University of California, Berkeley. Our estimates of where the people shaping AI studied and where their careers led, from education records in cited biographies.
  4. Google has the highest combined h-index among AI organizationsIncludes University of California, Berkeley. Our estimates of combined and median h-index at AI organizations, using current affiliations and OpenAlex records.

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 Collin Burns yet. Send us a free-licensed photo.

  1. Collin Burns collinpburns.com. Checked 19 October 2026.
  2. Measuring Massive Multitask Language Understanding arxiv.org. Checked 19 October 2026.
  3. Aligning AI With Shared Human Values arxiv.org. Checked 19 October 2026.
  4. Measuring Mathematical Problem Solving With the MATH Dataset arxiv.org. Checked 19 October 2026.
  5. Collin Burns On Discovering Latent Knowledge In Language Models Without Supervision theinsideview.ai. Checked 19 October 2026.
  6. Discovering Latent Knowledge in Language Models Without Supervision arxiv.org. Checked 19 October 2026.
  7. Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision proceedings.mlr.press. Checked 19 October 2026.
  8. Unsupervised Elicitation of Language Models arxiv.org. Checked 19 October 2026.
  9. Unsupervised Elicitation alignment.anthropic.com. Checked 19 October 2026.
  10. Chunky Post-Training: Data Driven Failures of Generalization arxiv.org. Checked 19 October 2026.
  11. IBB Econ & Tech INsights — June 2026 centerforchinaanalysis.asiasociety.org. Checked 19 October 2026.
  12. Trump's latest AI czar has already resigned techcrunch.com. Checked 19 October 2026.