Pan tiewen

A single-rollout RLVR method achieves faster training and more stable gradients by using a trajectory's own future as a hindsight context, eliminating the need for group sampling and reference solutions.