I am a french Math PhD advised by Loucas Pillaud-Vivien and Francis Bach at Sierra, Inria Paris. My research interest are in understanding deep-learning systems through mathematical/theoretic tools, in particular during training and deployment: How does a model train ? What computation does a trained model do ? Here is a list of resources illustrating these questions :

• Singular Learning Theory : wild maths that link singular algebra to bayesian machine learning.
• Mechanistic interpretability : one can try to recover algorithms learned by foundational models to understand various tasks like memorization / translation / elementary algebra.
• Superposition, Grokking and associated phenomenons.
• Implicit biases : when neural networks interpolate data, many interpolating solutions exists, and I am interested in understanding the properties of the one found by gradient descent-like algorithms.

AI Safety work

I am quite worried about the risk that powerful AIs may pose to society, and I am willing to oriente ma carrier differently to adress (if possible) some of these issues.

In particular, I was a fellow MATS 4.0 student, ML4G bootcamps teacher, and I teach the Turing seminar at Master MVA.

Feel free to contact me for information or simply discuss AI Safety in France :)

About me

I used to volunteer for important, meaningful and fun associations. I was involved in EffiSciences, which I cofounded, and where I helped develop and promote the idea of Recherche Impliquée (RI), (close to Impactful Research), which says that research can and should try to have a positive impact on the world. EffiSciences has a lot of different projects on which I contributed since 2022 in RI, biosecurity and safe AIs. We wrote a report on impactful research that you can find here (2025).

Other stuff: I like to read and sometime post on LessWrong. My favorite sport is {biking, running, weight-lifting, bjj, climbing}. I love to discuss philosophy, with some favorite topics being : Absurdism (Camus), Consciousness, Language (Wittgenstein), Moral.

My CV.

Publications

• Convergence of Shallow ReLU Networks on Weakly Interacting Data, Léo Dana, Loucas Pillaud-Vivien, Francis Bach, accepted at NeurIPS 2025. (Work done at Sierra, Inria Paris.)
• Rapport Recherche Impliquée, Jérémy Andréoletti, Antoine Sérandour , Léo Dana , Léo Karoubi (Work done at EffiSciences).
• Memorization in Attention-only Transformers, Léo Dana, Muni Sreenivas Pydi, Yann Chevaleyre, accepted at AISTATS 2025. (Work done at Lamsade, Paris-Dauphine University.)