Why Diffusion Models Don't Memorize

NeurIPS 2025 Best Paper (one of four from 21,575 submissions), from ENS Paris/PSL's LPENS physics-of-learning group (Bonnaire, Urfin, Biroli): the memorization timescale grows linearly with dataset size while the generalization timescale stays constant, so early stopping is a structural necessity rather than a heuristic. Directly load-bearing for how the field reasons about diffusion memorization — part of the group's multi-year statistical-physics program on diffusion dynamics.