question about RL training implementation detail

[Patiskey]

I have a question: why is the retrieval step fixed at 2 for reinforcement learning training on multi-hop task datasets, instead of allowing free retrieval to converge?