About
I am a second-year PhD student at École Normale Supérieure – PSL in Paris, supervised by Giulio Biroli (ENS) and Marc Mézard (Bocconi University). I am based at the Centre de Sciences des Données and the Laboratoire de Physique de l’ENS (LPENS).
My research aims to build a theoretical understanding of diffusion models, currently the dominant approach to generative AI, using tools from statistical physics.
I was very glad that our work on why diffusion models do not memorize their training data received one of the four Best Paper Awards at NeurIPS 2025.
News
- Our paper Weighting Schedules Govern What and When Score-Based Generative Models Learn from Multimodal Data, with Jérémie Klinger, Giulio Biroli and Marylou Gabrié, was accepted as a poster at the NeurIPS 2026 workshop Geometric Distributional Deep Learning. Come see our poster in Paris!
- Looking forward to presenting Double Descent and Malign Overfitting in Diffusion Models at the GdR IASIS day on generative models at ENS Lyon.
- Two new preprints out: First Learn, Then Memorize: The Spectral Bias of Diffusion Models and Weighting Schedules Govern What and When Score-Based Generative Models Learn from Multimodal Data.
Contact
Email. raphael.urfin at phys.ens.psl.eu
Office. Centre de Sciences des Données, 45 rue d'Ulm, 75005 Paris.
Profiles. Google Scholar · OpenReview · GitHub · LinkedIn