Checkpoint Handoffs Split an Agentic RL Gain Into Where the Agent Arrives and What It Does There
arXiv 2609.19636 (submitted 17 Sep 2026) points out that an agent in a closed loop writes its own inputs, so an SFT and an RL checkpoint are scored from different states even on identical tasks, and restricting comparison to states both policies reach selects on an outcome, which in their data flips the sign of the effect. The proposed protocol, checkpoint handoff, clones a state one released checkpoint reached and hands it to another with no retraining, splitting an endpoint gain into REACH (arriving at a state the environment confirms is a fixed number of actions from success) and SOLVE (finishing from an identical cloned state). Across two benchmarks and two independently released pipelines, the reacher-by-solver interaction is positive in all five conditions, and on ALFWorld the SFT solver never succeeds where the RL solver fails.
↳ Follow the thread