Alireza Abdollahpoorrostam

M.Sc. Researcher, EPFL · Lausanne, Switzerland

prof_pic.jpg

Chair of Risk Analytics & Optimization

Signal Processing Lab (LTS4)

EPFL, Lausanne, Switzerland

I am a graduate researcher at EPFL, working at the intersection of the Chair of Risk Analytics and Optimization (RAO) with Prof. Daniel Kuhn and the Signal Processing Laboratory (LTS4) with Prof. Pascal Frossard.

I am drawn to the mathematical foundations of deep learning — why neural networks fail under adversarial perturbations and distribution shift, and what principled algorithms follow. My work spans adversarial robustness, generalization, and post-training methods (model merging, model editing) for foundation models, with recent work on the optimal-transport geometry of adversarial training, distributionally robust optimization, and efficient LLM pretraining through optimizer-aware data selection.

I aim to derive rigorous guarantees and turn them into deployable algorithms. My research has appeared at NeurIPS, ICML, and ICLR.

Before EPFL, I completed a B.Sc. in Electrical Engineering & Computer Science at Amirkabir University of Technology (Tehran Polytechnic).


Research interests

  • Adversarial robustness and certified defenses
  • Distributionally robust optimization & optimal transport
  • LLM post-training and pre-training
  • Model merging, model soups, and weight-space methods

Recent research

  • Brenier Meets Adversarial Training — accepted at NeurIPS 2026. We use optimal-transport geometry to build structured adversaries for robust learning (with Ehsan Sharifian, Buse Şen, Marco Cuturi, and Daniel Kuhn).
  • OptiSelect — a new preprint with Simin Fan and Martin Jaggi on optimizer-aware data selection and training recipes for efficient LLM pretraining.
  • MonoSoup — a hyperparameter-free, data-free post-hoc method that recovers Model-Soup-level OOD accuracy from a single fine-tuned checkpoint, reducing ensembling cost from $\mathcal{O}(K)$ to $\mathcal{O}(1)$ (ICML 2026).

news

Nov 9, 2026 Brenier Meets Adversarial Training: Optimal Transport Geometry for Robust Learning has been accepted at NeurIPS 2026 — with Ehsan Sharifian, Buse Şen, Marco Cuturi, and Daniel Kuhn.
Oct 2, 2026 New preprint with Simin Fan and Martin Jaggi: OptiSelect: How does the Optimizer Shape Data Curriculum? — optimizer-aware data selection and training recipes for efficient LLM pretraining.
May 13, 2026 New blog post: The Centaur’s Gambit — how AI is elevating scientific research, just as engines revolutionized chess.
May 1, 2026 Our paper Model Soups Need Only One Ingredient has been accepted to ICML 2026 — with N. Dimitriadis, A. Hazimeh, and P. Frossard.
Jan 22, 2026 A General Framework for Black-Box Attacks under Cost Asymmetry (with M. Salmani and S.-M. Moosavi-Dezfooli) accepted to ICLR 2026.

latest posts

selected publications

  1. NeurIPS
    Brenier Meets Adversarial Training: Optimal Transport Geometry for Robust Learning
    Alireza Abdollahpoorrostam, Ehsan Sharifian, Buse Şen, and 2 more authors
    In Advances in Neural Information Processing Systems (NeurIPS), 2026
    Accepted
  2. Preprint
    OptiSelect: How does the Optimizer Shape Data Curriculum?
    Simin Fan, Alireza Abdollahpoorrostam, and Martin Jaggi
    arXiv preprint, 2026
  3. ICML
    Model Soups Need Only One Ingredient
    Alireza Abdollahpoorrostam, Nikolaos Dimitriadis, Adam Hazimeh, and 1 more author
    In International Conference on Machine Learning (ICML), 2026
  4. ICLR
    A General Framework for Black-Box Attacks under Cost Asymmetry
    Mohammadreza Salmani, Alireza Abdollahpoorrostam, and Seyed-Mohsen Moosavi-Dezfooli
    In International Conference on Learning Representations (ICLR), 2026
  5. NeurIPS
    SuperDeepFool: A New Fast and Accurate Minimal Adversarial Attack
    Alireza Abdollahpoorrostam, Mahed Abroshan, and Seyed-Mohsen Moosavi-Dezfooli
    In Advances in Neural Information Processing Systems (NeurIPS), 2024