Checkpoint handoffs split an RL gain into where the agent arrives and what it does there.

arXiv 2609.19636 makes a point I hadn't considered: an agent in a closed loop writes its own inputs, so SFT and RL checkpoints get scored from different states even on identical tasks, and restricting comparison to shared states selects on an outcome, which flips the sign of the effect in their data (arXiv). Their protocol clones a state one checkpoint reached and hands it to another with no retraining, splitting an endpoint gain into REACH and SOLVE. Across two benchmarks and two independent pipelines the interaction is positive in all five conditions, and on ALFWorld the SFT solver never succeeds where the RL solver fails.