Arya Akhavan

Welcome! I’m Arya Akhavan, a postdoctoral researcher at the University of Oxford, working with Patrick Rebeschini. My current research explores how memory can enhance reasoning and support self-improvement in large language models and reinforcement learning agents. This interest grew out of my visit to Csaba Szepesvári at the University of Alberta in March 2026. I am particularly interested in how agents can accumulate, retrieve, and build on past experience to improve their reasoning and decision-making without updating their model parameters.

Together with Csaba and Patrick, I investigate the capabilities of memory-augmented systems, with a focus on enhancing reasoning and advancing AI safety. Within this research direction, I also collaborate with Murat A. Erdogdu on how memory can help make AI systems safer and more reliable. This collaboration began during my visit to the Vector Institute.

Previously, I spent a year as a postdoctoral fellow at École Polytechnique, working at the intersection of online learning and optimization. I completed my PhD at Institut Polytechnique de Paris and the Istituto Italiano di Tecnologia (IIT) in Genoa, where I was fortunate to be advised by Alexandre B. Tsybakov and Massimiliano Pontil.

Work Experience

  • University of Oxford

    Postdoctoral Researcher, September 2024 – Present

    Working with Patrick Rebeschini

  • University of Alberta

    Visiting Researcher, March 2026 – May 2026

    Working with Csaba Szepesvári

  • Vector Institute and University of Toronto

    Visiting Researcher, May 2026

    Working with Murat A. Erdogdu

  • École Polytechnique de Paris

    Postdoctoral Researcher, March 2023 – July 2024

    Working with Karim Lounici and El Mahdi El Mhamdi

Education

  • PhD in Theoretical Machine Learning

    ENSAE IP, Paris, and Istituto Italiano di Tecnologia-Genova (IIT), September 2019 – February 2023

    Supervisors: Alexandre B. Tsybakov and Massimiliano Pontil

  • Master M2 MASH Learning and Applications

    Paris Dauphine University and École Normale Supérieure Paris-Saclay,   Fall 2017 – December 2018

  • First year of Master in Mathematics

    Sharif University of Technology,   2016 – 2017

  • B.Sc. in Mathematics

    Department of Mathematics, Statistics and Computer Sciences, University of Tehran,   2011 – 2016

Arya Akhavan by the lake in St James's Park, London

Work in Progress

  • External memory for reliable long-horizon reasoning and state tracking

    A. Akhavan, D. Janz, P. Rebeschini, and C. Szepesvári

    Manuscript in preparation.

  • Learning when to ask: Adaptive expert queries for imitation learning

    A. Masoumian, A. Akhavan, and C. Szepesvári

    Manuscript in preparation.

  • Memory-augmented AI safety: Learning when and how to intervene in language model reasoning

    A. Akhavan, M. A. Erdogdu, P. Rebeschini, and C. Szepesvári

    Ongoing research project investigating when and how memory-based interventions can improve the safety and reliability of language-model reasoning.

  • From past solutions to future plans: Reasoning across tasks with growing memory

    A. Akhavan, D. Gomez, and C. Szepesvári

    Ongoing research project studying how agents can reuse and expand memories of prior solutions to plan effectively across tasks.

Publications and Preprints