PeopleofSuperintelligence

Researcher

Micah Goldblum

Micah Goldblum is a machine learning researcher who studies why deep neural networks work and how to make them safe and secure. They earned a PhD in mathematics at the University of Maryland in 2020, then spent 2021 to 2024 as a postdoctoral researcher at New York University with Yann LeCun and Andrew Gordon Wilson. A 2022 paper on the marginal likelihood, written with Sanae Lotfi, Pavel Izmailov, Gregory Benton and Andrew Gordon Wilson, won an ICML Outstanding Paper award, and in 2024 they coauthored the LiveBench benchmark for language models. In May 2026, they spent four weeks at École Polytechnique as an Alliance Visiting Professor. They are an assistant professor at Columbia University in electrical engineering and computer science, a post they took up in July 2024, and a co-founder of the Columbia Core AI Lab.

Timeline

  1. 2014–2020

    A Maryland mathematics PhD

    From September 2014 to May 2020, they work toward a PhD in mathematics at the University of Maryland with Tom Goldstein and Wojciech Czaja. Their research there covers meta-learning and robust computer vision.

  2. December 2020

    Dataset security for machine learning

    In December 2020, they post Dataset Security for Machine Learning: Data Poisoning, Backdoor Attacks, and Defenses as first author, with Dimitris Tsipras, Chulin Xie, Xinyun Chen, Avi Schwarzschild, Dawn Song, Aleksander Madry, Bo Li and Tom Goldstein. The paper categorizes dataset vulnerabilities and defenses and develops a unified taxonomy of poisoning and backdoor threat models.

  3. September 2021–June 2024

    Postdoc at New York University

    From September 2021 to June 2024, they are a postdoctoral researcher at New York University, working with Yann LeCun and Andrew Gordon Wilson.

  4. 2022

    An ICML Outstanding Paper award

    In 2022, they share an Outstanding Paper award at ICML for Bayesian Model Selection, the Marginal Likelihood, and Generalization, written with Sanae Lotfi, Pavel Izmailov, Gregory Benton and Andrew Gordon Wilson. The paper shows that the marginal likelihood can be negatively correlated with generalization and offers a conditional marginal likelihood as a partial remedy.

  5. June 2024

    LiveBench for language models

    In June 2024, they and 17 coauthors, among them Yann LeCun and Tom Goldstein, release LiveBench, a benchmark for large language models whose questions are updated frequently from recent information sources and scored automatically against objective ground-truth values. The paper is an ICLR 2025 Spotlight.

  6. 2025–2026

    Small batches and Zebra-CoT

    In 2025, they coauthor Small Batch Size Training for Language Models with Martin Marek, Sanae Lotfi, Aditya Somasundaram and Andrew Gordon Wilson for NeurIPS, which shows that small batch sizes down to one train language models stably and make vanilla SGD work. In 2026, they coauthor Zebra-CoT at ICLR, a dataset of 182,384 interleaved text-image reasoning traces across 18 domains.

  7. May 2026

    Visiting professor at Polytechnique

    In May 2026, they spend four weeks at École Polytechnique's Center for Applied Mathematics (CMAP) as an Alliance Visiting Professor. They work with El Mahdi El Mhamdi on two projects, one on AI safety and one on training algorithms for large language models.

  8. As of October 2026

    Assistant professor at Columbia

    They join Columbia University in July 2024 and are an assistant professor there in electrical engineering and computer science. They co-found the Columbia Core AI Lab (CAIL), and their research spans training, architecture and inference strategies for large-scale models, AI safety and agents.

Career

  1. Since 2024Columbia UniversityAssistant Professor, Co-founder
  2. 2026École PolytechniqueAlliance Visiting Professor
  3. 2021–2024New York UniversityPostdoctoral Researcher

Education

  1. 2020University of MarylandPhD

From the biography.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Columbia University. 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. Nonprofit roles are 2.3 times as common in AI safetyIncludes Columbia University. We estimate where people focused on safety, alignment and evaluation work. Nonprofit roles appear for 25.1% of safety people with known affiliations, compared with 11.0% in other areas.
  3. Universities had a net outflow of 2,068 moves in AI careersIncludes Columbia University. Our estimates of career flows and median h-index across universities and companies in AI, including the role of joint appointments and earlier academic jobs.
  4. People shaping AI work at a median of 3 organizations over a career: 3 for university starters and institute and lab starters, versus 2 for company startersIncludes Columbia University. Across 6,853 people shaping AI with a dated work role, we estimate a median of 3 parent organizations over a career (mean 3.6), comparing role groups, starting sectors and starting decades from the 1950s to 2020–2026.

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 Micah Goldblum yet. Send us a free-licensed photo.

  1. Micah Goldblum goldblum.github.io. Checked 21 October 2026.
  2. Micah Goldblum | alphaXiv alphaxiv.org. Checked 21 October 2026.
  3. Group - UMD Department of Computer Science cs.umd.edu. Checked 21 October 2026.
  4. Dataset Security for Machine Learning: Data Poisoning, Backdoor Attacks, and Defenses arxiv.org. Checked 21 October 2026.
  5. ICML 2022 Awards icml.cc. Checked 21 October 2026.
  6. Bayesian Model Selection, the Marginal Likelihood, and Generalization arxiv.org. Checked 21 October 2026.
  7. LiveBench: A Challenging, Contamination-Limited LLM Benchmark arxiv.org. Checked 21 October 2026.
  8. Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation Is Wasteful arxiv.org. Checked 21 October 2026.
  9. Strengthening international research in IA with an Alliance Visiting Professorship polytechnique.edu. Checked 21 October 2026.