generalfactneutralPolicy optimization can achieve data-dependent regret bounds with first-order, second-order, and path-length guarantees even under unknown transition kernelsMachine Learning (Statistics)27 Jul 2026http://arxiv.org/abs/2606.31769v1