benchmarksfactneutralOnline recalibration can be achieved in approximately epsilon^-3 rounds for Lipschitz proper losses, which is optimalMachine Learning (Statistics)28 Jul 2026http://arxiv.org/abs/2607.19692v1