Tag
This paper introduces SARA, a sequential adaptive rollout allocation method for RLVR that abandons saturated groups early and reallocates the budget, achieving comparable accuracy with 22% fewer rollouts than dynamic sampling and up to 67% savings when combined.