Contact me at
matteo.negri1@cyu.fr
I am a tenure track professor (CPJ) at the Laboratoire de Physique Théorique et Modélisation, CY Cergy Paris Université.
Previously:
Post-Doc at Università di Roma Sapienza
PhD in Physics at Politecnico di Torino
Master's Degree in Physics at Università degli studi di Milano
How does creativity emerge from the memorzation task? Can this explain when generative models truly generate, and when they just reproduce training data? My research addresses these questions using tools from the statistical physics of disordered systems. The central object is the associative memory: a system that stores data as basins of attraction in an energy landscape and retrieves them under partial or noisy inputs. We showed that modern architectures are associative memories in disguise, and that their capacity to generalise to unseen data emerges from a sharp phase transition controlled by training set size: on one side of the transition, the model memorizes; on the other, it creates new samples.
Jun 2026 Invited speaker at Mathematical Foundations of AI and Statistical Physics of Deep Learning III
Apr 2026 Awarded the CY Initiative EMERGENCE grant (50k€)
Apr 2026 New preprint Language Diffusion Models are Associative Memories (arXiv)
Apr 2026 Poster at ICLR on neural scaling laws + Workshop on Associative Memories
Dec 2025 New Job! Joined LPTM, CY Cergy Paris Université as Chaire de Professeur Junior
I firmly believe in the social role of science and scientists, but I am invested in science popularization also because I really enjoy it. I’ve been involved in initiatives for high-school students organized by University of Rome Sapienza and in some history-of-physics lectures for adults organized by a book store. I participated to the 2024 FameLab competition, which also includes training sessions for the participants. You can find more here (mostly in italian).