Reach or Solve? Attributing Agentic RL Gains with Checkpoint Handoffs

Xuan Liu (Shanghai Jiao Tong University) and Jingbin Qian (Rice University) introduce checkpoint handoff, which clones a state one released checkpoint reached and hands it to another, splitting an agentic RL endpoint gain into REACH and SOLVE.
Ask this paper
Endpoint success confounds two changes. Because an agent in a closed loop writes its own later inputs, SFT and RL checkpoints are scored from different states even on identical tasks.
Filtering to shared states makes it worse. Restricting comparison to states both policies reach selects on an outcome, and in their data it flips the sign of the effect.
The reacher-by-solver interaction is positive in all five conditions. Across two benchmarks and two independently released pipelines, an RL history is worth more to an RL solver than the same history is to an SFT solver.
The solver effect is 2 to 5 times larger on RL-produced states. On ALFWorld, RL improves both terms and the SFT solver never succeeds where the RL solver fails; with scale the gain moves from the solver to the reacher.
The protocol needs no retraining. Handoff requires only that one checkpoint's history be replayable under another, so long-horizon evaluations can report arrival and completion beside endpoint success.
Abstract
Reinforcement learning now trains language-model agents that act over dozens of steps in live environments. The gains are large, and they are read as better decision-making. An agent in a closed loop writes its own inputs. Each observation follows from its own earlier actions, so the states it meets late in an episode are partly of its own making. An SFT checkpoint and an RL checkpoint are then scored from different states, even on identical tasks. Endpoint success mixes two changes: where the agent arrives, and what it does once it is there. Restricting the comparison to states both policies reach does not separate them. That restriction selects on an outcome, and in our data it flips the sign of the effect. We introduce checkpoint handoff, an evaluation protocol that clones a state one released checkpoint reached and hands it to another, with no retraining. Crossing a reacher role and a solver role over SFT and RL splits an endpoint gain into REACH and SOLVE. REACH is how often a policy arrives at a state the environment confirms is a fixed number of actions from success. SOLVE is how often it finishes from an identical cloned state. Across two benchmarks and two independently released pipelines, the reacher by solver interaction is positive in all five conditions. An RL history is worth more to an RL solver than the same history is to an SFT solver. On ALFWorld, RL improves both terms, and the SFT solver never succeeds where the RL solver fails. Independent REACH and SOLVE gaps predict the aggregate interaction. Handoff asks only that one checkpoint's history can be replayed under another, so long-horizon evaluation can report arrival and completion beside endpoint success.