PeopleofSuperintelligence

Researcher

Raphaël Urfin

Raphaël Urfin is a PhD student in physics and machine learning at École Normale Supérieure – PSL who uses statistical physics to build a theory of diffusion models, the dominant approach to generative AI. From 2021, they studied physics and mathematics at ENS, interned in 2023 in the Soft Matter Group at Cambridge, and worked in 2024 and 2025 with Giulio Biroli and Marc Mézard on diffusion models. With Tony Bonnaire, Biroli and Mézard, they wrote "Why Diffusion Models Don't Memorize," which found two training timescales and won one of four Best Paper Awards at NeurIPS 2025. The PhD began in September 2025, the paper appeared in the Journal of Statistical Mechanics in August 2026, and three new preprints on diffusion models followed in September 2026. As of October 2026, they are a second-year PhD student based at the Centre de Sciences des Données and LPENS.

Timeline

  1. 2021–2024

    Physics and mathematics at ENS

    They enter ENS-PSL in Paris in 2021 and take a bachelor in physics and mathematics with first class honors. In 2024, they complete the M2 ICFP theoretical physics track there, with classes in advanced statistical physics, disordered systems and machine learning.

  2. 2023

    Ecology research at Cambridge

    In 2023, they do a research internship with Camille Scalliet in the Soft Matter Group at the University of Cambridge's DAMTP. The project uses cavity-method and random-matrix tools from disordered systems to study collective behavior in ecological systems.

  3. April 2024–May 2025

    Internships on diffusion models

    From April 2024 to January 2025, they do a research internship with Giulio Biroli at the Centre de Sciences des Données of ENS-PSL on the effect of implicit regularization in diffusion models. From February to May 2025, they work with Marc Mézard in the Department of Computing Sciences at Bocconi University on the memorization-generalization transition in diffusion models, which leads to a paper accepted at NeurIPS 2025.

  4. 2025

    Best Paper at NeurIPS 2025

    With Tony Bonnaire, Giulio Biroli and Marc Mézard, they write "Why Diffusion Models Don't Memorize: The Role of Implicit Dynamical Regularization in Training." The paper finds an early timescale at which models generate good samples and a later memorization timescale that grows linearly with training-set size. NeurIPS 2025 lists it among its four Best Papers and accepts it as an oral.

  5. September 2025

    Starting a PhD at ENS

    They start a PhD at ENS-PSL in September 2025, working with Giulio Biroli and Marc Mézard.

  6. August 2026

    Journal version in JSTAT

    The NeurIPS paper appears in August 2026 in the Journal of Statistical Mechanics: Theory and Experiment, in the special issue Machine Learning 2026. The published version marks Tony Bonnaire and Urfin as equal contributors.

  7. September 2026

    Three diffusion-model preprints

    Three preprints on diffusion models appear in September 2026. "First Learn, Then Memorize: The Spectral Bias of Diffusion Models" and "Double Descent and Malign Overfitting in Diffusion Models" come with Tony Bonnaire, Giulio Biroli and Marc Mézard. "Weighting Schedules Govern What and When Score-Based Generative Models Learn from Multimodal Data" comes with Jérémie Klinger, Giulio Biroli and Marylou Gabrié.

  8. As of October 2026

    PhD student at ENS-PSL

    They are a second-year PhD student at ENS-PSL, based at the Centre de Sciences des Données and the Laboratoire de Physique de l'ENS (LPENS), and build a theoretical understanding of diffusion models with tools from statistical physics. In October 2026, the weighting-schedules paper with Jérémie Klinger, Giulio Biroli and Marylou Gabrié is accepted as a poster at the NeurIPS 2026 workshop Geometric Distributional Deep Learning.

Career

  1. Since 2024Université Paris Sciences et LettresResearch Intern
  2. 2025Bocconi University
  3. 2023Department of Applied Mathematics and Theoretical Physics, University of CambridgeResearch Intern

Education

  1. From 2025Université Paris Sciences et LettresPhD
  2. 2024Université Paris Sciences et LettresBachelor's and M2

From the biography.

  1. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes Université Paris Sciences et Lettres. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.
  2. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionIncludes Université Paris Sciences et Lettres. We estimate where the people shaping AI studied and where their careers led. Massachusetts Institute of Technology leads with 467 former students, including 291 recorded doctorates.
  3. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Université Paris Sciences et Lettres. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.
  4. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Université Paris Sciences et Lettres. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Raphaël Urfin yet. Send us a free-licensed photo.

  1. CV — Raphaël Urfin urfin-raphael.github.io. Checked 22 October 2026.
  2. [2505.17638] Why Diffusion Models Don't Memorize: The Role of Implicit Dynamical Regularization in Training arxiv.org. Checked 22 October 2026.
  3. Announcing the NeurIPS 2025 Best Paper Awards – NeurIPS Blog blog.neurips.cc. Checked 22 October 2026.
  4. Marc Mézard Receives NeurIPS 2025 Award | Bocconi University unibocconi.it. Checked 22 October 2026.
  5. News — Raphaël Urfin urfin-raphael.github.io. Checked 22 October 2026.
  6. Why diffusion models do not memorize: the role of implicit dynamical regularization in training - Journal of Statistical Mechanics: Theory and Experiment iopscience.iop.org. Checked 22 October 2026.
  7. Raphaël Urfin urfin-raphael.github.io. Checked 22 October 2026.