generalfactneutralPAC-Bayesian generalization certificates become tighter when the geometry-induced prior has smaller KL divergence from the posteriorMachine Learning (Statistics)28 Jul 2026http://arxiv.org/abs/2607.18422v1