PeopleofSuperintelligence

Researcher

Mahdi Haghifam

Mahdi Haghifam is a machine learning researcher whose theory work covers generalization, memorization and privacy in learning algorithms, and whose recent interests include distillation and inference-time algorithms for language models. After a B.Sc. and an M.Sc. at Sharif University of Technology in Tehran, he completed a PhD in machine learning at the University of Toronto and the Vector Institute, advised by Daniel Roy. In 2024, he coauthored “Information Complexity of Stochastic Convex Optimization: Applications to Generalization, Memorization, and Tracing”, which received a Best Paper award at ICML 2024. He then worked as a Distinguished Postdoctoral Fellow at Northeastern University's Khoury College with Jonathan Ullman, joined the Toyota Technological Institute at Chicago as a research assistant professor in 2025, and was a Research Fellow at the Simons Institute in Spring 2026. His 2026 affiliation is Google Research.

Badges

Timeline

  1. By 2019

    A PhD at Toronto and Vector

    After a B.Sc. and an M.Sc. at Sharif University of Technology in Tehran, he works toward a PhD in machine learning at the University of Toronto and the Vector Institute, advised by Daniel M. Roy. In 2019, he interns at ServiceNow Research in Toronto with Gintare Karolina Dziugaite, where he proposes data-dependent generalization estimates for noisy SGD and SGLD.

  2. 2020–2021

    Conditional mutual information bounds

    In 2020, he coauthors “Sharpened Generalization Bounds based on Conditional Mutual Information and an Application to Noisy, Iterative Algorithms” with Jeffrey Negrea, Ashish Khisti, Daniel M. Roy and Gintare Karolina Dziugaite, which shows that bounds built on conditional mutual information are tighter than those built on unconditional mutual information and applies them to Langevin dynamics. A 2021 paper with Dziugaite, Shay Moran and Roy, “Towards a Unified Information-Theoretic Framework for Generalization”, investigates whether the conditional mutual information framework can unify generalization bounds in the realizable setting.

  3. 2022–2023

    A Google Brain/DeepMind internship

    From September 2022 to December 2022, he is a research intern at Google Brain/DeepMind in Mountain View, working with Thomas Steinke. In 2023, he coauthors “Why Is Public Pretraining Necessary for Private Model Training?” with Arun Ganesh, Milad Nasr, Sewoong Oh, Thomas Steinke, Om Thakkar, Abhradeep Thakurta and Lun Wang, which proposes that training must first select a good basin in the loss landscape, a task that is harder with private data.

  4. 2024

    An ICML Best Paper

    In 2024, he coauthors “Information Complexity of Stochastic Convex Optimization: Applications to Generalization, Memorization, and Tracing” with Idan Attias, Gintare Karolina Dziugaite, Roi Livni and Daniel Roy, and it receives a Best Paper award at ICML 2024. The paper characterizes the tradeoff between the accuracy of a learning algorithm and the conditional mutual information it reveals about its training data in stochastic convex optimization.

  5. By 2025

    Postdoc at Northeastern's Khoury College

    By 2025, he is a Distinguished Postdoctoral Fellow at Northeastern University's Khoury College of Computer Sciences, advised by Jonathan Ullman, and he also works with Adam Smith. His research there centers on the theoretical foundations of trustworthy machine learning, with an emphasis on privacy and generalization.

  6. 2025–2026

    Research assistant professor at TTIC

    In 2025, he joins the Toyota Technological Institute at Chicago (TTIC) as a research assistant professor. In August 2025, he submits “The Sample Complexity of Membership Inference and Privacy Auditing” with Adam Smith and Jonathan Ullman, which shows for the first time that a membership-inference attacker sometimes needs many more samples than the training algorithm uses, and it is a FOCS 2026 paper.

  7. 2026

    Research Fellow at the Simons Institute

    In Spring 2026, he is a Research Fellow at the Simons Institute for the Theory of Computing at UC Berkeley, taking part in the Federated and Collaborative Learning program. His recent interests include problems motivated by language models, such as distillation and inference-time algorithms.

  8. Since 2026

    Google Research affiliation

    His TTIC appointment runs from 2025 to 2026, and his affiliation in 2026 is Google Research. His research interests are machine learning, trustworthy AI, data privacy and information theory.

Career

  1. Since 2026Google Research
  2. 2026University of California, BerkeleyResearch Fellow
  3. Until 2025Northeastern UniversityDistinguished Postdoctoral Fellow
  4. 2025–2026Toyota Technological Institute at ChicagoResearch Assistant Professor
  5. 2022GoogleResearch Intern
  6. 2019ServiceNowResearch Intern

Education

  1. Sharif University of TechnologyB.Sc.
  2. University of TorontoPhD
  3. Vector InstitutePhD

From the biography.

  1. 16% of AI career moves in the 2020s went from a university to a companyIncludes Mahdi Haghifam. 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.
  2. 23% of people shaping AI who left a university for a company later returned to a universityIncludes Mahdi Haghifam. We estimate that 468 of the 2,011 people shaping AI who left a university for a company (23%) later returned to a university after a median of 5 years, and 163 of 829 (20%) made the reverse round trip.
  3. Google had the most people shaping AI every year from 2014 to 2025Includes Google. Our estimates of which organizations employed the most people shaping AI each year from 1955 to 2025, tracing the shift from university labs to companies.
  4. Google has the highest combined h-index among AI organizationsIncludes Google. We estimate that Google holds the highest combined h-index at 8,016, while universities account for 73% of the deduplicated total across AI organizations.

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 Mahdi Haghifam yet. Send us a free-licensed photo.

  1. Mahdi Haghifam mhaghifam.github.io. Checked 21 October 2026.
  2. [2004.12983] Sharpened Generalization Bounds based on Conditional Mutual Information and an Application to Noisy, Iterative Algorithms arxiv.org. Checked 21 October 2026.
  3. [2111.05275] Towards a Unified Information-Theoretic Framework for Generalization arxiv.org. Checked 21 October 2026.
  4. [2302.09483] Why Is Public Pretraining Necessary for Private Model Training? arxiv.org. Checked 21 October 2026.
  5. ICML 2024 Awards icml.cc. Checked 21 October 2026.
  6. Khoury PhD students, postdocs make the jump to professor in 2025 - Khoury College of Computer Sciences khoury.northeastern.edu. Checked 21 October 2026.
  7. TTIC Faculty Alumni ttic.edu. Checked 21 October 2026.
  8. [2508.19458] The Sample Complexity of Membership Inference and Privacy Auditing arxiv.org. Checked 21 October 2026.
  9. Mahdi Haghifam simons.berkeley.edu. Checked 21 October 2026.