Execution is often used as terminal feedback: it scores a generated program, but the resulting state is not reused for further learning. Learning after Execution treats the execution result as an intermediate state, enabling an additional learning interaction for the same policy. This creates a credit‑assignment problem because the two actions share the final objective yet play different roles in the trajectory. Action‑specific credit assigns credit to the initial action from the final trajectory outcome and assigns corrective credit to the improvement action relative to the intermediate state. Image‑to‑TikZ serves as an executable visual testbed. On 945 held‑out examples, learning after execution strengthens the shared policy under single‑call inference and yields a post‑execution corrective capability. Single‑call render success reaches 92.70%, compared with 88.25% for continued single‑stage GRPO after the same number of additional training iterations. This corrective capability cannot be reproduced by continued single‑stage training and depends on the intermediate state. Alternative credit definitions induce different group‑relative preferences over the same trajectories. Together, the results suggest that executable tasks contain useful learning structure beyond a terminal reward. Code is available at https://github.com/tianyufengcn/Learning-after-Execution.
Review