VINE

Researchers have identified that training instability in reinforcement learning with flow-matching policies is caused by the vanilla sampling strategy, not iterative action generation. They introduced…