2026-09-22 · America/Los_Angeles · 论文 · #9

One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents

Repository-level software engineering (SWE) comprises heterogeneous task categories, whose progress under pooled agentic reinforcement learning can be uneven: gains in some categories coincide with regressions in others, while aggregate resolution obscures these changes. Motivated by this category see-saw, we develop a category-aware expert-training and policy-integration framework. Executable task construction and SWE Labeler, an evidence-grounded multi-axis labeling system, organize the training pools. Initial category-specific RL improves average training success while leaving uneven…

热度 52.4 / 100；排名与评分保留该期记录。