Aymane El Firdoussi

Aymane El Firdoussi

PhD Student

Cambridge, United Kingdom

About

I’m a PhD student at the University of Cambridge working (mostly) on Generative Modelling and Reinforcement Learning.

I completed the MVA master’s program (Mathématiques, Vision et Apprentissage) at Institut Polytechnique de Paris together with an engineering degree in applied mathematics and computer science at Télécom Paris. I am also fortunate to have worked with amazing researchers at Google DeepMind, EPFL and the Technology Innovation Institute.

I am usually open to collaborations on the topics listed below. Feel free to reach out via email if you want to work together on exciting projects.

Publications

Check my Google Scholar profile for an updated list of my publications.

Research interests

  1. Diffusion Generative Modeling

    Diffusion Language Models, Diffusion on discrete data, Sampling unnormalised densities with Diffusion Samplers, Stochastic Optimal Control.

  2. Reinforcement Learning for Generative Models

    Post-training LLMs and Diffusion Models with RL, Preference Optimization.

  3. High-Dimensional Statistics

    Random Matrix Theory for Machine Learning, Exact asymptotics of learning when dimension and sample size grow together.

Software

  • Research Paper Reading Tool

    A local app to keep track of the papers read. Papers are organised by topic, and each one gets a TLDR, a summary and notes written in LaTeX, edited side by side with the PDF.

  • DECOR

    The official code repository of the paper: 'The Privacy Power of Correlated Noise in Decentralized Learning', published at ICML 2024, where we propose a SOTA algorithm for Differentially Private Decentralized Learning.

  • Alpha-LoRA

    The official code repository of the paper: 'Effective Fine-Tuning via Base Model Rescaling'. We propose a theoretically proven improvement to Low Rank Adaptation by acting on the weights of the Base Model.

  • Synthetic Data

    The official code repository of the paper: 'Maximizing the Potential of Synthetic Data: Insights from Random Matrix Theory', published at ICLR 2025. We leveraged Random Matrix Theory to propose a theoretical framework to understand the effect of synthetic data on a model's training performance.