generalfactbullishGRAS remains effective even for non-differentiable rewardsit remains effective even for non-differentiable rewardsMachine Learning (Statistics)29 Aug 2026http://arxiv.org/abs/2608.26585v1