Hossein Mohebbi
Portrait of Hossein Mohebbi Portrait of Hossein Mohebbi

Hossein Mohebbi

I’m a machine learning researcher working on reinforcement learning, causal inference and robust learning. Lately I’ve been focused on computer-use agents and RL environments for expert domains.

  • Research at stealth lab
  • BCS, MMath, University of Waterloo
  • Previously Amazon, Google and the Vector Institute

About

Most of my work circles one question: how do we get more out of the models and data we already have? That has meant reinforcement learning as a decision layer over pretrained models, counterfactual augmentation when labelled data are scarce, and now computer-use agents and multi-agent systems.

I did my MMath in Computer Science at the University of Waterloo with Pascal Poupart, writing a thesis on counterfactual data augmentation whose core method appeared at ICML 2026. During the degree I was a researcher at the Vector Institute and a Student Researcher at Google, where I worked with Suraj Kothawade on reinforcement learning for image generation.

I also have years of production engineering behind me: I was a software engineer at Amazon, and during my undergraduate degree I did co-op terms at BitGo, AWS, StackAdapt, ApplyBoard, IBM and High Q.

News

  1. Presented CRDA at ICML 2026 in Seoul, and at the RIKEN–Vector–NAIRL Workshop.

  2. CRDA is on arXiv, with a project page and code: pip install crda.

  3. Talk on CRDA at Vector Institute Community Day, Toronto.

  4. Image-POSER is on arXiv.

  5. Talk on Image-POSER at Vector Institute Research Day, Toronto.

  6. Started research on computer-use agents at an AI lab in stealth.

Research

Current · since 2025

Data and environments for agents

Agents learn from whatever signal we can give them, and good signal is scarce. At a lab still in stealth, I work on two sources of it. The first is reinforcement-learning environments for expert workflows: real tasks turn into a sandboxed environment with an automatic verifier, hardened against reward hacking so it rejects the shortcuts a model would use to score without solving the task. Many agents can then work in parallel, with the verifier refereeing every attempt.

The second is human demonstrations of computer use. Agents that act through a mouse and keyboard are held back less by architecture than by data, because there is no large, open record of how people actually use diverse computer applications. I work on that data, the screenshot–action trajectories built from it, and the open agents and benchmarks it makes possible. The models, benchmarks and datasets are intended for open release.

Papers — Library Search papers RLVR is the way Saved 012345

Actions

  1. a0click(226, 58)
  2. a1type("RLVR is the way")
  3. a2key("Enter")
  4. a3scroll(0, +2)
  5. a4click(300, 154)
  6. a5click(514, 58)

τ = (s0, a0, …, a5, s6)

Figure 1. A computer-use trajectory τ = (s0, a0, s1, a1, …). At each step the agent observes a screenshot st and emits a low-level action at: click, type, key or scroll. Human trajectories like this one are the raw material for training and evaluating agents.

2025 · Google

Reinforcement learning over generative experts

No single image model is best at everything. Image-POSER casts image generation and editing as a Markov decision process. A long, compositional prompt is broken into subtasks, and at each step a learned policy chooses which expert to call from a registry of pretrained text-to-image and image-to-image models. A vision–language critic inspects every intermediate image and supplies step-wise reward, so the policy learns when to generate, when to edit, and with which model.

It outperforms frontier baselines on alignment, fidelity and aesthetics, and human evaluators consistently prefer its outputs.

Paper Image-POSER: Reflective RL for Multi-Expert Image Generation and Editing Preprint

Prompt “Lighthouse on a cliff at dusk, a red sailboat in the bay, in watercolour.” Subtasks 1 Lighthouse at dusk 2 Add a red sailboat 3 Watercolour style rt π Policy Text-to-image Image-to-image T2I · 1 T2I · 2 I2I · edit I2I · inpaint I2I · style Image st Vision–language critic
Figure 2. Image-POSER. The policy π routes each subtask to one expert from a registry of pretrained text-to-image (T2I) and image-to-image (I2I) models. A vision–language critic scores the intermediate image st and returns reward rt. (Schematic; expert names are placeholders.)

2023–2026 · Waterloo & Vector Institute

Counterfactual data augmentation

Once a regressor has captured the systematic part of the data, what remains — the residual — behaves like noise, and that noise should not change when certain features do. CRDA perturbs statistically screened features and re-attaches each observation’s residual, producing counterfactual training samples without collecting anything new.

The method is model-agnostic and comes with a finite-sample consistency result for the screening filter and a causal robustness bound. Across benchmark datasets it reduces average MSE by 22.9% for MLP regressors and 6.4% for XGBoost. It is the basis of my master’s thesis and appeared at ICML 2026.

Paper Counterfactual Residual Data Augmentation for Regression ICML 2026

x y f(x) f + ε ε
Figure 3. Residual reuse in one dimension. A base regressor f explains the systematic part of each observation (x, y), leaving a residual ε = y − f(x). Perturbing x and re-attaching ε gives counterfactual samples (x′, f(x′) + ε) that lie on the residual-shifted curve. Hover or tap to pick an observation.

Publications

Conference papers

  1. ICML 2026

    Counterfactual Residual Data Augmentation for Regression

    Hossein Mohebbi, Oliver Schulte, Ke Li, Pascal Poupart

    43rd International Conference on Machine Learning (ICML 2026), Seoul, South Korea

    A model-agnostic augmentation method for tabular regression that treats residuals as invariant noise under perturbations of statistically screened features, with consistency and robustness guarantees.

    PaperProject pageCode

Preprints

  1. Preprint

    Image-POSER: Reflective RL for Multi-Expert Image Generation and Editing

    Hossein Mohebbi†, Mohammed Abdulrahman†, Yanting Miao, Pascal Poupart, Suraj Kothawade

    Under review · † equal contribution

    Reinforcement learning orchestrates a registry of pretrained image generation and editing experts, supervised step by step by a vision–language critic.

    Paper

Thesis

  1. Thesis

    Counterfactual Data Augmentation for Regression

    Hossein Mohebbi

    MMath thesis, University of Waterloo, 2026. Advisor: Pascal Poupart

    UWSpace

Experience

Research

  1. Aug 2025 – Present

    Founding Researcher · AI lab in stealth

    Leading the design of reinforcement-learning environments for expert domains, with sandboxed tasks and verifiers hardened against reward hacking; research on open computer-use agents, benchmarks and the human data behind them.

  2. Jan – Dec 2025

    Student Researcher · Google

    Built Image-POSER in a sponsored collaboration with Pascal Poupart (Waterloo) and Suraj Kothawade (Google); studied the efficiency and inference cost of unified image-generation models.

  3. Sep 2023 – Dec 2025

    Machine Learning Researcher · Vector Institute

    Developed CRDA and its supporting theory; ran distributed experiments on Alliance Canada HPC clusters.

  4. Sep – Dec 2019

    Research Assistant · University of Waterloo

    With Mohammad Kohandel: statistical modelling of how cancer stem cells respond to chemotherapy and immunotherapy.

Industry

  1. Jul 2023 – Feb 2025

    Software Development Engineer · Amazon

    At RDS, built automation for the database fleet that rolls out engine updates and maintenance safely across AWS regions, with monitoring and alarms that catch problems before customers do. At Buy with Prime, worked on the APIs and integrations that connect merchants’ stores and e-commerce platforms to Buy with Prime, keeping catalog, inventory and orders in sync, and scaled services for peak traffic like Prime Day.

  2. 2019 – 2023

    Software Engineer · co-op terms

    BitGo
    Ethereum staking infrastructure; unstaking support for the Shapella upgrade.
    Jan – Apr 2023
    Amazon
    AWS region-build automation; cut RDS admin-database builds from two days to 40 minutes.
    May – Aug 2022
    StackAdapt
    Backend services for roles and permissions; front-end tooling.
    Jan – Apr 2022
    ApplyBoard
    An eligibility API matching students with academic programs.
    Sep 2020 – Apr 2021
    IBM
    Extensions for QRadar, IBM’s security analytics platform.
    Jan – Apr 2020
    High Q
    Control and monitoring software for quantum-computing hardware.
    May – Aug 2019

Education

  1. 2023 – 2025

    Master of Mathematics (MMath), Computer Science · University of Waterloo

    Thesis: Counterfactual Data Augmentation for Regression · Advisor: Pascal Poupart

  2. 2018 – 2023

    Bachelor of Computer Science (BCS), Honours, Co-operative Program · University of Waterloo

    Human–Computer Interaction Option · Dean’s Honours

Talks

  1. 2026

    Counterfactual Residual Data Augmentation for Regression

    • PosterICML 2026, SeoulJul 2026
    • PosterRIKEN–Vector–NAIRL Workshop, SeoulJul 2026
    • TalkVector Institute Community Day, TorontoMay 2026
  2. 2025

    Image-POSER: Reflective RL for Multi-Expert Image Generation and Editing

    • TalkVector Institute Research Day, TorontoOct 2025

Teaching

  1. Fall 2024

    CS 114: Principles of Computing for Science · University of Waterloo

    Teaching Assistant

  2. Spring 2024

    CS 231: Algorithmic Problem Solving · University of Waterloo

    Teaching Assistant

  3. May 2024

    Reinforcement Learning Bootcamp · Vector Institute

    Teaching Assistant. Taught foundational RL and ran hands-on workshops for ten industry partner teams, guiding their proof-of-concept projects.

  4. Fall 2023

    CS 458: Computer Security and Privacy · University of Waterloo

    Guest Lecturer. Deep RL for smart-contract security: a DQN agent that finds fund-leaking (“prodigal”) vulnerabilities in Ethereum contracts.

Honours

  1. 2025

    President’s Graduate Scholarship · University of Waterloo

  2. 2025

    Ontario Graduate Scholarship

  3. 2020

    President’s Research Award · University of Waterloo

  4. 2019

    President’s Scholarship of Distinction · University of Waterloo

  5. 2018 – 2023

    Dean’s Honours List · University of Waterloo

Contact

Email is the best way to reach me. I’m always happy to talk about computer-use agents, reinforcement learning, and possible collaborations.

Google Scholar GitHub LinkedIn