What does "the terminal node’s confidence score achieved from rollouts" mean in paper?

[tongyx361]

We compute each trajectory’s final score by multiplying its reward with the terminal node’s confidence score achieved from rollouts.

Why does the terminal node still have rollouts?

[Thunderbeee]

In MCTS, a "rollout" involves simulating a sequence of actions from a given node to a terminal state to estimate the potential outcome or reward of that path. However, once a terminal node is reached, it represents the end of a reasoning path, and no further rollouts are performed from that point. Instead, the terminal node's confidence score is derived from the rollouts conducted during the generation of that trajectory and majority voting at the terminal nodes. This confidence score reflects the model's assessment of the correctness or quality of the reasoning path leading to that terminal state.