Splitting repository-level SWE tasks into category experts and distilling them back into one model lifts Pro-618 mean resolution to 58.04%
Addressing a "category see-saw" in repository-level software engineering, where pooled agentic reinforcement learning improves some task categories while regressing others and aggregate resolution hides the change, the work builds a category-aware expert-training and policy-integration framework: SWE Labeler groups tasks by repository domain into service/data/security (A), user-facing applications (B), and systems/tooling/runtimes (C); Agentic-miniRL with a Refresh–Repair–Expand loop trains same-origin category experts; and label-routed multi-teacher on-policy distillation (MOPD) consolidates them into one deployable student, which reaches 58.04% mean resolution on Pro-618 (+5.39 points over base) and 59.00% on SWE-bench Multilingual (+2.