Chan Li and Nigel Goldenfeld published an accepted paper in Physical Review Letters on November 2, 2026. They use dynamical mean field theory to show that double descent behavior is a phase transition in the training process. The study calculates critical exponents and identifies a breakdown of the fluctuation-dissipation theorem. This breakdown is associated with broken ergodicity and relates to the network's ability to generalize accurately.
Read the full story at journals.aps.org.
Photo: amazing_designs / Pixabay
