Tag
This paper introduces the checkpoint handoff protocol to attribute gains in agentic reinforcement learning by separating 'Reach' (arriving at useful states) and 'Solve' (solving from those states), showing RL improvements stem from both components.