Marc Mézard Receives NeurIPS 2025 Award
Marc Mézard (Invernizzi Chair in Computer Science at Bocconi) is among the authors of the paper “Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training” together with Tony Bonnaire, Raphaël Urfin e Giulio Biroli (École Normale Supérieure), selected as one of the award-winning contributions at NeurIPS 2025, one of the world’s leading conferences in artificial intelligence and machine learning. Held annually, NeurIPS gathers thousands of researchers and is regarded as a key venue where foundational advances in the field are first presented. The recognition of Mézard’s work highlights a study that addresses one of the most debated issues in generative AI: understanding how diffusion models produce new content rather than replicating their training data. This is an extraordinary achievement, considering the numbers: 21,575 submissions in total, 5,290 accepted, and only 77 invited to give an oral presentation. The paper by Mézard and colleagues thus joins the elite, in the top 0.36% of contributions submitted.
The study demonstrates that the training of diffusion models unfolds along two distinct timescales. During the initial phase, marked by a stable characteristic time, the model learns to generate high-quality samples. Only later can tendencies toward memorization emerge, associated with a much longer timescale that grows linearly with dataset size. When training sets are large, this separation creates an extended window in which the model generalizes effectively and generates original content without reproducing real data.
The NeurIPS award committee noted that the paper stands out for offering a clear theoretical explanation to an important and unresolved question regarding the non-memorization behavior of diffusion models. The study identifies a general learning mechanism, described as implicit dynamical regularization, which arises naturally from the training dynamics and may prove relevant to other machine-learning architectures. The committee also emphasized the value of the interdisciplinary approach adopted by the authors, who apply concepts from statistical physics to modern neural-network theory, contributing to a theoretically significant advance for the field.
As Mézard observes: "Understanding the mechanisms that enable diffusion models to avoid memorization is not only a theoretical challenge, but also a fundamental step toward building more reliable and robust artificial intelligence systems. This work demonstrates that it is the training dynamics themselves that act as a form of regularization."
