In software‑engineering (SWE) tasks, fixing a repository‑level bug requires an agent to perform three stages: localization, reproduction, and patching. Traditional reinforcement‑learning approaches train solely from the final test result of a patch, offering no guidance for the intermediate steps. On‑policy self‑distillation (OPSD) adds token‑level supervision by exposing the full reference patch at each step, yet it still lacks stage‑specific signals for localization or reproduction.
SWE‑Ladder addresses this by converting the verified repair of each training instance into stage‑specific evidence: changed files and line numbers for localization, fail‑to‑pass test pairs for reproduction, and the diff for patching. A rule‑based router assigns every step of an end‑to‑end rollout to a stage according to the invoked tool; a frozen teacher sees only the evidence matching that stage. The stages are then distilled in the order of repair while GRPO optimizes the entire trajectory.
Experiments with Qwen3‑4B‑Instruct‑2507 on the SWE‑bench under the R2E‑Gym scaffold show that SWE‑Ladder raises pass@1 from 9.6% (baseline) to 16.4%, outperforming GRPO (10.4%) and OPSD with the full patch (12.4%). Under the OpenHands scaffold, pass@1 improves from 13.0% to 17.6%. Reversing the stage order in R2E‑Gym drops pass@1 to 11.4%, highlighting the importance of the correct stage sequence.
The results demonstrate that progressive privileged self‑distillation with stage‑wise evidence can substantially boost an agent’s ability to localize, reproduce, and patch software bugs. Review