2026_03_31

A new preprint out: A Noise Sensitivity Exponent Controls Large Statistical-to-Computational Gaps in Single- and Multi-Index Models. We characterize wide statistical-to-computational gaps for high noise in single-index models and for large separable multi-index models, as a function of a ``noise sensitivity exponent’’, a simple property of the activation function. This is another joint work with Leonardo Defilippis (ENS), Florent Krzakala (EPFL), and Bruno Loureiro (ENS).