Curriculum vitae
↓ Download PDFEducation
PhD in Physics and Machine Learning — ENS-PSL, Paris
Subject: Generative diffusion and statistical physics · LPENS & Centre de Sciences des Données
Supervisors: Giulio Biroli (ENS) and Marc Mézard (Bocconi University)
2025 – 2028 (exp.)
ENS Diploma — ENS-PSL, Paris
Interdisciplinary diploma validating my studies at ENS.
2021 – 2025
M2 ICFP, Theoretical Physics Track — ENS-PSL, Paris
First Class Honors, 17.38/20.
Relevant classes: Advanced Statistical Physics, Disordered Systems, Machine Learning.
2023 – 2024
M1 ICFP — ENS-PSL, Paris
First Class Honors, 17.8/20.
Relevant classes: Phase Transitions, Introduction to Quantum Field Theory, General Relativity.
2022 – 2023
Bachelor in Physics and Mathematics — ENS-PSL, Paris
First Class Honors, 17.38/20 (Physics) and 16.36/20 (Mathematics).
2021 – 2022
Publications
* denotes equal contribution.
Conference proceedings
Tony Bonnaire*, Raphaël Urfin*, Giulio Biroli, Marc Mézard.
Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training.
NeurIPS 2025 — Best Paper Award (4 papers out of 20,000+ submissions) and Oral.
Advances in Neural Information Processing Systems 38, 141266–141286. Proceedings
Selected for the JSTAT special issue Machine Learning 2026: J. Stat. Mech. (2026) 084007.
2025
Preprints
Raphaël Urfin, Tony Bonnaire, Giulio Biroli, Marc Mézard.
First Learn, Then Memorize: The Spectral Bias of Diffusion Models.
arXiv:2609.35377. arXiv
2026
Jérémie Klinger, Raphaël Urfin, Giulio Biroli, Marylou Gabrié.
Weighting Schedules Govern What and When Score-Based Generative Models Learn from Multimodal Data.
arXiv:2609.35322. arXiv
2026
Raphaël Urfin*, Tony Bonnaire*, Giulio Biroli, Marc Mézard.
Double Descent and Malign Overfitting in Diffusion Models.
arXiv:2609.26392. arXiv
2026
See the publications page for figures and BibTeX.
Conferences, workshops & seminars
Invited talks
Generative Memory Lab — Donders Institute, Radboud University, Netherlands (online)
Invited seminar
Sep 2026
LemanTh 2026 — Lausanne, Switzerland
Workshop
Apr 2026
Vitelli’s Group Meeting — online
Group meeting (non-reciprocal p-spins)
Mar 2026
Scattering Club — Data Science for Astrophysics group, LPENS (online)
Invited seminar (1 hour)
Feb 2026
Voxel51 “Best of NeurIPS 2025” — online
Invited seminar (30 min) · Video
Jan 2026
Centre de Sciences des Données Internal Seminar — Paris, France
Invited seminar (30 min)
Dec 2025
Machine Learning & Signal Processing @ ENS Lyon — Lyon, France
Invited seminar (1 hour)
Nov 2025
Contributed talks
GdR IASIS Day on Generative Models (diffusion & flow matching) — ENS Lyon, France
Workshop (25-min talk)
Oct 2026
Journées de Physique Statistique 2026 — Paris, France
Conference (4-min flash talk)
Jan 2026
EurIPS 2025 — Copenhagen, Denmark
Conference (poster + 13-min oral)
Dec 2025
NeurIPS in Paris 2025 — Paris, France
Workshop (poster + 10-min oral)
Nov 2025
StatPhys 29 — Florence, Italy
Conference (15-min talk)
Jul 2025
Youth in High Dimensions — Trieste, Italy
Workshop (15-min talk) · Video
Jul 2025
Journées de Physique Statistique 2025 — Paris, France
Conference (4-min flash talk)
Jan 2025
Summer schools
Beg Rohu Summer School of Statistical Physics — Beg Rohu, France
- 2026 Edition (Optimization and Control of Nonequilibrium Systems): Lectures by international researchers (e.g., G. Peyré, G. Rotskoff...). Presented a poster on Memorization/Generalization in diffusion models.
- 2025 Edition (Learning with Machines, Physics and Mind): Lectures on Machine Learning and Statistical Physics by international researchers (e.g., Yann LeCun, Julia Kempe, Stéphane Mallat, Marc Mézard...). Presented a poster on Memorization/Generalization in diffusion models.
Jun 2025 &
Jun 2026
Jun 2026
Complex and Glassy Systems — Cargèse, France
Lectures on statistical physics and interdisciplinary applications by international researchers (e.g., Marc Mézard, Eric Vanden-Eijnden, Valentina Ros, Guy Bunin).
Jul 2024
Awards & grants
Best Paper Award, NeurIPS 2025
One of four among 20,000+ submissions.
2025
CDSN — competitive PhD scholarship for ENS students
2025
Academic activities
Teaching
TA, Phase Transitions (M1 ICFP, ENS-PSL)
16 hours, taught in English with Giulio Biroli.
Fall 2026
TA, Mathematics tutoring for first-year physics students at ENS
36 hours, taught in French with Amir-Kian Kashani-Poor.
Fall 2025 &
Fall 2026
Fall 2026
TA, Machine Learning (M2 ICFP, ENS-PSL)
11 hours, taught in English with Marc Lelarge and Leonardo Defilippis.
Spring 2026
TA, Stochastic Processes for Physics (M1 ICFP, ENS-PSL)
16 hours, taught in English with Marylou Gabrié.
Fall 2025
“Khôlles” in Mathematics and Physics — Lycée Stanislas, Paris
Preparation for the oral exams of the French “Grandes Écoles”, Classes Préparatoires PC.
2022 – 2024
Student supervision
Sacha Gaucher — Master's research internship, ESPCI Paris
Co-supervised with Giulio Biroli. Subject: the RePaint method (inpainting with diffusion models).
Apr – Jul 2026
Reviewing
- Workshops: EurIPS 2025 (Principles of Generative Modeling — PriGM), ICLR 2026 (DeLTa), ICLR 2026 (Sci4DL).
- Journals: Transactions on Machine Learning Research (TMLR).
Skills
- Languages. French (native), English (fluent), Italian (fluent).
- Software. Python (PyTorch), Matlab, LaTeX.
Research experience
Statistical Physics & Diffusion Models — Bocconi University, Milan, Italy
Research internship with Marc Mézard (Dept. of Computing Sciences). The memorization/generalization transition in diffusion models. Resulted in a publication accepted at NeurIPS 2025.
Feb – May 2025
Statistical Physics & Diffusion Models — ENS-PSL, Paris
Research internship with Giulio Biroli (Centre de Sciences des Données). The effect of implicit regularization in diffusion models.
Apr 2024 – Jan 2025
Emergent Behaviors in Large Ecosystems — University of Cambridge, UK
Research internship with Camille Scalliet (Soft Matter Group, DAMTP). Using tools from disordered systems (cavity method, random matrix theory...) to understand emergent collective behaviors in ecological systems.
Feb – Jul 2023