generalfactneutralThe geometry-induced prior approximates the ideal but inadmissible posterior-matched prior, which would minimize the KL term by depending on training dataMachine Learning (Statistics)28 Jul 2026http://arxiv.org/abs/2607.18422v1