PeopleofSuperintelligence

Researcher

Rob Brekelmans

Rob Brekelmans is a machine-learning researcher who works on probabilistic inference for generative models, from language models to diffusion models. After early papers on total correlation explanation in 2018 and annealed importance sampling in 2020, he earned a PhD at the University of Southern California in 2022, having worked with Greg Ver Steeg and Aram Galstyan. He coauthored Action Matching with Kirill Neklyudov, Daniel Severo and Alireza Makhzani in 2023, and won an ICML 2024 Best Paper award with Stephen Zhao, Alireza Makhzani and Roger Grosse for twisted Sequential Monte Carlo in language models. A 2025 paper on Feynman-Kac Correctors brought Sequential Monte Carlo resampling to pretrained diffusion models. He is a postdoctoral fellow at the Vector Institute in Toronto, working with Alireza Makhzani and Roger Grosse.

Timeline

  1. 2018

    Auto-encoding total correlation explanation

    In 2018, he coauthors "Auto-Encoding Total Correlation Explanation" with Shuyang Gao, Greg Ver Steeg and Aram Galstyan. The paper introduces a variational lower bound to total correlation explanation (CorEx) and finds that it is equivalent to the variational autoencoder bound under certain conditions.

  2. 2020

    Annealed importance sampling with q-paths

    In 2020, he and six coauthors, including Vaden Masrani, Frank Wood and Frank Nielsen, explore annealed importance sampling with q-paths, a family that includes the geometric path as a special case. The paper carries a Best Paper Award at the NeurIPS Workshop on Deep Learning through Information Geometry.

  3. 2022

    PhD at USC and DeepMind

    He graduates with a PhD from the University of Southern California in 2022, having worked with Greg Ver Steeg and Aram Galstyan, and interns with the AI Safety Analysis team at DeepMind alongside Pedro Ortega and Tim Genewein. In 2022, he coauthors "Your Policy Regularizer is Secretly an Adversary" with Genewein, Ortega, Shane Legg and others, showing that the robustness of policy regularization comes from hedging against worst-case reward perturbations.

  4. 2023

    Action Matching at ICML

    He coauthors Action Matching with Kirill Neklyudov, Daniel Severo and Alireza Makhzani, presented at the 40th International Conference on Machine Learning in 2023. The method learns the dynamics of a system from independent samples of its time evolution, with a training objective that avoids back-propagation through differential equations or optimal transport solvers.

  5. 2024

    ICML 2024 Best Paper

    At ICML 2024, "Probabilistic Inference in Language Models via Twisted Sequential Monte Carlo", by Stephen Zhao, Rob Brekelmans, Alireza Makhzani and Roger Grosse, receives a Best Paper award. The method uses learned twist functions to estimate the expected future value of the potential at each timestep, which focuses inference-time computation on promising partial sequences.

  6. 2025

    Feynman-Kac Correctors for diffusion

    The 42nd International Conference on Machine Learning in 2025 publishes "Feynman-Kac Correctors in Diffusion: Annealing, Guidance, and Product of Experts", which he coauthors with Marta Skreta, Alexander Tong, Kirill Neklyudov, Arnaud Doucet and others. The method derives a weighted simulation scheme for sampling from annealed, geometric-averaged or product distributions of pretrained score-based models, using Sequential Monte Carlo resampling.

  7. As of October 2026

    Postdoc at the Vector Institute

    He works as a postdoctoral fellow at the Vector Institute in Toronto, with Alireza Makhzani and Roger Grosse.

Career

  1. NowVector InstitutePostdoctoral Fellow
  2. 2022Google DeepMindIntern

Education

  1. 2022University of Southern CaliforniaPhD

From the biography.

  1. Of 943 people shaping AI who worked at Google, 446 work there nowIncludes Rob Brekelmans. Our estimates of where the 943 people shaping AI who worked at Google and Google DeepMind work now: 446 at Google and 452 across 278 other organizations.
  2. Nonprofit roles are 2.3 times as common in AI safetyIncludes Vector Institute. 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 (31%) and startups and other companies (25%) employ the largest shares of 2020–2026 entrantsIncludes Vector Institute. Among 819 people whose first recorded role started in 2020–2026 and who have a known current organization, 60% work in commercial sectors (25% at startups and other companies, 23% at frontier labs and 13% at big tech, with 9 in two), and 31% work at universities, compared with 66% for earlier entrants.
  4. The United States employs an estimated 48% of people shaping AIIncludes Vector Institute. We estimate where the people shaping AI work today, how universities, companies and research institutes split across countries, and how country shares have changed since the 1960s.

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 Rob Brekelmans yet. Send us a free-licensed photo.

  1. [1802.05822] Auto-Encoding Total Correlation Explanation arxiv.org. Checked 22 October 2026.
  2. [2012.07823] Annealed Importance Sampling with q-Paths arxiv.org. Checked 22 October 2026.
  3. Rob Brekelmans brekelma.github.io. Checked 22 October 2026.
  4. [2203.12592] Your Policy Regularizer is Secretly an Adversary arxiv.org. Checked 22 October 2026.
  5. Action Matching: Learning Stochastic Dynamics from Samples proceedings.mlr.press. Checked 22 October 2026.
  6. ICML 2024 Awards icml.cc. Checked 22 October 2026.
  7. [2404.17546] Probabilistic Inference in Language Models via Twisted Sequential Monte Carlo arxiv.org. Checked 22 October 2026.
  8. Feynman-Kac Correctors in Diffusion: Annealing, Guidance, and Product of Experts proceedings.mlr.press. Checked 22 October 2026.