A scientific breakthrough reveals why generative AI learns so well

The rapid rise of AI has given rise to a new generation of models capable of producing images, sounds, or videos of impressive realism. Among them, generative models stand out because, by learning from numerous examples, they are able to create content that is often indistinguishable from real data. But behind this feat lies a fundamental challenge: how do these systems manage to invent new data (images, sounds, videos, etc.)—that is, to generalize—rather than simply memorizing and then repeating exactly what they have “learned”? 
 
Using an interdisciplinary approach combining statistical physics, computer science, and numerical experiments, Tony Bonnaire and his collaborators have made a key discovery regarding the learning process of diffusion models: the quantitative identification of two distinct and predictable time scales—an initial generalization phase independent of the training data, followed much later by a memorization phase dependent on the size of the dataset.
The team shows that the memorization time decreases as the amount of training data increases, thereby explaining why generative AIs based on diffusion models remain in a phase where they create new data.
 
By demonstrating that the observed performance of diffusion models and their practical success are based on a provable and measurable mechanism that naturally delays overfitting, the work of Tony Bonnaire and his collaborators offers a deep and actionable understanding of the mechanisms governing modern generative AI. It thus establishes a new foundation for the scientific analysis of generalization in artificial intelligence, as underscored by the award of the NeurIPS 2025 Best Paper Award. This is an exceptional honor, as only 4 papers were recognized out of the 21,575 submitted (of which 5,290 were accepted) to this most prestigious conference in the field of AI.