Repository-level software engineering (SWE) spans heterogeneous task categories, and pooled agentic reinforcement learning can produce uneven progress: gains in some categories coincide with regressions in others, a phenomenon the authors call a "category see-saw." Aggregate resolution metrics obscure these per-category shifts.
To address this, the researchers introduce a category-aware expert-training and policy-integration framework. It combines executable task construction with SWE Labeler, an evidence-grounded multi-axis labeling system that organizes training pools. Initial category-specific RL improves average training success but leaves uneven instance-level progress, motivating explicit consolidation of successful behavior and policy-adaptive task selection.
Same-origin category experts alternate long-horizon Agentic-miniRL with Refresh-Repair-Expand (RRE): the updated policy refreshes instance mastery, reuses its own verified successful trajectories for Repair SFT, and reselects tasks for further RL. Label-routed multi-teacher on-policy distillation (MOPD) then consolidates experts into one deployable student, using ReLU-gated reward extrapolation to keep only each teacher's improving direction over the reference. Notably, expert training and policy integration require no external model to supply solution trajectories or action targets.
Evaluation covers Pooled RL and Balanced RL, expert development, and single-model integration, measured through aggregate and per-category resolution, the minimum category lift over each joint-RL baseline, and expert-gain recovery. The final MOPD policy achieves mean resolution of 58.04% on Pro-618 and 59.00% on SWE-bench Multilingual, improving over the base model by 5.39 and 2.78 percentage points, respectively.
The model is released on Hugging Face as Logics-SWE-Qwen3.6-27B, alongside the Logics-SWE-Env-2.5K dataset and the open-source SWE-Labeler repository. Execution-environment container images referenced by the dataset are available through GitHub Container Registry (GHCR), supporting reproducible training and evaluation.