Ph.D. Student, Department of Computer Science, University of Toronto, mousavi [@] cs (dot) toronto (dot) edu

About me

I am a computer science Ph.D. student at the University of Toronto and Vector Institute, advised by Murat A. Erdogdu. Before that, I received my B.Sc. in computer engineering from Sharif University of Technology. During Fall 2023, I visited Lénaïc Chizat at the EPFL Institute of Mathematics. During Summer 2025, I interned at Apple ML Research with Marco Cuturi. I spent the spring of 2026 as a visiting researcher in Servicenow AI Research. During Summer 2026, I will participate in the MATS Fellowship, working with Roger Grosse.

Research Interests

I am broadly interested in developing a principled understanding of deep learning, both by analyzing current algorithms and by discovering new ones. Some topics I am focusing on:

• Theoretical Foundations of Modern AI: How do different stages of training (pre-training/SFT/RL) and deploying (parallel/sequential test-time scaling) current large models interact with each other? What are their statistical and computational complexities and tradeoffs? How to provably predict the scaling of different architectures?

• Generative Modeling: How to make flow models and diffusions more efficient?

• Training Dynamics of Neural Nets: How do first-order optimization algorithms explore the loss landscape and learn useful features in neural nets?

Selected Publications

• Post-Training with Policy Gradients: Optimality and the Base Model Barrier.
  Alireza Mousavi-Hosseini, Murat A. Erdogdu.
  To appear in ICML 2026, International Conference on Machine Learning. (Spotlight)

• Flow Matching with Semidiscrete Couplings.
  Alireza Mousavi-Hosseini*, Stephen Y. Zhang*, Michal Klein, Marco Cuturi.
  ICLR 2026, International Conference on Learning Representations.

• When Do Transformers Outperform Feedforward and Recurrent Networks? A Statistical Perspective.
  Alireza Mousavi-Hosseini, Clayton Sanford, Denny Wu, Murat A. Erdogdu.
  NeurIPS 2025, Advances in Neural Information Processing Systems.
  Oral Presentation, ICML workshop on High-Dimensional Learning Dynamics, 2025.

• Gradient-Based Feature Learning under Structured Data.
  Alireza Mousavi-Hosseini, Denny Wu, Taiji Suzuki, Murat A. Erdogdu.
  NeurIPS 2023, Advances in Neural Information Processing Systems.

• Towards a Complete Analysis of Langevin Monte Carlo: Beyond Poincaré Inequality.
  Alireza Mousavi-Hosseini*, Tyler Farghly*, Ye He, Krishnakumar Balasubramanian, Murat A. Erdogdu.
  COLT 2023, Annual Conference on Learning Theory.

• Neural Networks Efficiently Learn Low-Dimensional Representations with SGD.
  Alireza Mousavi-Hosseini, Sejun Park, Manuela Girotti, Ioannis Mitliagkas, Murat A. Erdogdu.
  ICLR 2023, International Conference on Learning Representations. (Spotlight)

Miscellaneous

• Toronto is a pretty cool city if you're interested in the performing arts in general, and musical theatre in particular.

• UofT has an amazing theatre: Hart House Theatre