PeopleofSuperintelligence

Researcher

Tomer Ullman

Also known as Tomer D. Ullman, Tomer David Ullman

Tomer Ullman is a cognitive scientist whose research connects models of human reasoning with AI systems that learn and reason like people. He earned a Ph.D. at MIT's Department of Brain and Cognitive Sciences in 2015 and coauthored work proposing causal models and intuitive theories as ingredients for machine learning. His research includes a benchmark for machine reasoning, tests of language and vision-language models, and studies of how alignment and inference-time controls affect model behavior. As of October 2026, he is the Morris Kahn Associate Professor of Psychology at Harvard, where he heads the Computation, Cognition, and Development lab and affiliates with the Kempner Institute.

Timeline

  1. 2015

    A thesis on intuitive theories

    In 2015, he completes a Ph.D. at MIT's Department of Brain and Cognitive Sciences with the dissertation “On the Nature and Origin of Intuitive Theories: Learning, Physics and Psychology.”

  2. 2016

    Building machines that learn

    In 2016, he coauthors “Building Machines That Learn and Think Like People,” arguing that AI should build causal models and ground learning in intuitive theories of physics and psychology.

  3. 2021

    The AGENT reasoning benchmark

    In 2021, he coauthors AGENT, a benchmark of procedurally generated 3D animations that probes core intuitive psychology in machine agents. It evaluates generalization and compares Bayesian inverse planning with a Theory of Mind neural network.

  4. 2022–2024

    A Jacobs Foundation fellowship

    From 2022 to 2024, he is a Jacobs Foundation Research Fellow and plans projects on play, learning, and social interaction. He studies common-sense reasoning through formal models of how children and adults learn new theories.

  5. 2023

    Theory of mind in language models

    In 2023, he tests large language models on theory-of-mind tasks and shows that small variations preserving Theory of Mind principles can reverse results from a reported success case.

  6. 2024

    Vision-language model illusions

    In a 2024 preprint, he tests vision-language models with “illusory-illusions,” cases near common visual illusions that should not elicit processing errors. Many systems mistakenly see these cases as illusions.

  7. 2025

    Alignment and conceptual diversity

    In 2025, he coauthors “One fish, two fish, but not the whole sea,” a study of conceptual diversity in synthetic language-model populations. Aligned models generally display less diversity than their instruction fine-tuned counterparts.

  8. 2025–2026

    Inference-time methods for AI

    His 2025 paper “Chain of Time” introduces cognitively inspired prompting that stabilizes physical simulation by image-generation models without fine-tuning. In a second paper submitted in 2025 and revised in 2026, he coauthors a Bayesian account of how prompts and activation steering affect language-model behavior.

  9. As of October 2026

    Professor and lab head at Harvard

    As of October 2026, he is the Morris Kahn Associate Professor of Psychology at Harvard and heads the Computation, Cognition, and Development lab. He is also an affiliate of the Kempner Institute for the Study of Natural and Artificial Intelligence, with a focus on intuitive theories and common-sense reasoning about physics and psychology.

Career

  1. NowHarvard UniversityMorris Kahn Associate Professor of Psychology, Lab Head

Education

  1. 2015Massachusetts Institute of TechnologyPhD

From the biography.

Sources and credits

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

  1. On the nature and origin of intuitive theories : learning, physics and psychology dspace.mit.edu. Checked 27 October 2026.
  2. Building Machines That Learn and Think Like People arxiv.org. Checked 27 October 2026.
  3. AGENT: A Benchmark for Core Psychological Reasoning proceedings.mlr.press. Checked 27 October 2026.
  4. Tomer Ullman - Jacobs Foundation jacobsfoundation.org. Checked 27 October 2026.
  5. Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks arxiv.org. Checked 27 October 2026.
  6. The Illusion-Illusion: Vision Language Models See Illusions Where There Are None arxiv.org. Checked 27 October 2026.
  7. One fish, two fish, but not the whole sea: Alignment reduces language models’ conceptual diversity aclanthology.org. Checked 27 October 2026.
  8. Chain of Time: In-Context Physical Simulation with Image Generation Models arxiv.org. Checked 27 October 2026.
  9. Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering arxiv.org. Checked 27 October 2026.
  10. Tomer D. Ullman | Research & Teaching tomerullman.org. Checked 27 October 2026.