M·F
Mahan Fathi, smiling, in a black t-shirt in front of a leafy tree

Mahan Fathi

ماهان فتحی

Research Scientist at NVIDIA

I work on post-training large language models at NVIDIA — reasoning through reinforcement learning with verifiable rewards (RLVR) and reward modeling — supervised by Olivier Delalleau.

Before that I did my master’s in computer science at Mila and Université de Montréal, advised by Pierre-Luc Bacon. Along the way I was a student researcher in Montréal under Ross Goroshin: first at Google Brain, on the MonRL team led by Marc G. Bellemare, then at Google DeepMind, on large language models and state-space models.

I’m proud of switching gears — pun intended. Coming from Mechanical Engineering at Sharif University of Technology, I decided not to settle for anything less than following my true passion, robotics; reinforcement learning just happened to be a recently popularized means to that end.

Across academia and industry, I’ve built large-scale research infrastructure for deep learning, in areas such as large language modeling, reinforcement learning and computer vision, and I’ve worked as a software engineer, data scientist and team leader on things like user-centric ML, high-throughput streaming microservices and image datasets at scale.

My research interests are pretty broad, but I find research that provokes questions far more valuable than work focused on providing answers.

Experience

  1. NVIDIA

    – Present

    Research Scientist, Post-training Canada · Remote

    Post-training of large language models: reasoning via RLVR and reward modeling. Supervised by Olivier Delalleau.

  2. Google DeepMind

    –

    Student Researcher, LLM Montréal, QC

    Large language modeling and state-space models. Supervised by Ross Goroshin.

  3. Google Brain

    –

    Student Researcher, Deep RL Montréal, QC

    Where sequence modeling meets offline and online reinforcement learning. A member of the MonRL team led by Marc G. Bellemare; supervised by Ross Goroshin.

  4. Mila Montreal Institute for Learning Algorithms

    –

    MSc Student Montréal, QC

    Advised by Pierre-Luc Bacon at Université de Montréal. Sample-efficient reinforcement learning, energy-based policy and imitation networks, and continuous, HiPPO-inspired sequence models in RL and control.

  5. Polytechnique Montréal

    –

    Teaching Assistant Montréal, QC

    For INF8953DE, Reinforcement Learning, taught by Sarath Chandar. Designed the majority of the assignments.

  6. Aalto University

    –

    Research Assistant Espoo, Finland

    Supervised by Perttu Hämäläinen. Built a framework that exposes MuJoCo dynamics steps as differentiable PyTorch operations, and used analytic policy gradients to control systems as complex as Humanoid.

Selected; I’m a firm believer in the one-page résumé. The full history is on LinkedIn.

Publications

  1. NeurIPS 2026

    Sparse Koopman Autoencoders Identify Local Dynamical Regimes in Multibasin Systems

    Aidan Li, Uday Kiran Reddy Tadipatri, Mahan Fathi, Sarath Chandar, Ross Goroshin

  2. ICML 2025

    In-Context Learning and Occam’s Razor

    Eric Elmoznino, Tom Marty, Tejas Kasetty, Leo Gagnon, Sarthak Mittal, Mahan Fathi, Dhanya Sridhar, Guillaume Lajoie

  3. ICLR 2024

    Course Correcting Koopman Representations

    Mahan Fathi, Clement Gehring, Jonathan Pilault, David Kanaa, Pierre-Luc Bacon, Ross Goroshin

  4. NeurIPS 2023

    Block-State Transformers

    Mahan Fathi*, Jonathan Pilault*, Orhan Firat, Christopher Pal, Pierre-Luc Bacon, Ross Goroshin

  5. MSc Thesis, 2024

    Beyond the Horizon: Improved Long-Range Sequence Modeling, from Dynamical Systems to Language

    Mahan Fathi, Université de Montréal

NVIDIA technical reports

  1. Technical Report, 2026

    Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI

    NVIDIA

  2. arXiv 2026

    Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

    NVIDIA

  3. arXiv 2026

    Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

    NVIDIA

  4. arXiv 2025

    NVIDIA Nemotron 3: Efficient and Open Intelligence

    NVIDIA

* Equal contribution. Everything is also on Google Scholar.

Education

  1. M.Sc. in Computer Science

    –

    Université de Montréal & Mila Dean’s List

  2. B.Sc. in Mechanical Engineering

    –

    Sharif University of Technology

  3. High School, NODET

    –

    National Organization for Development of Exceptional Talents Mathematics & Physics