Yingwei Ma, Yue Liu, Xinlong Yang, Yanhao Li, Kelin Fu, Yibo Miao, Yuchong Xie, Zhexu Wang, Shing-Chi Cheung

Researchers propose a new training paradigm for LLM coding agents that decomposes complex software engineering tasks into five atomic skills: code localization, code editing, unit-test generation, issue reproduction, and code review. Rather than optimizing agents on composite benchmarks like bug-fixing, which leads to task-specific overfitting, the team uses joint reinforcement learning to improve these foundational skills simultaneously. The approach shows 18.7% average performance gains across both atomic skills and unseen composite tasks, suggesting that mastery of atomic skills generalizes better than traditional task-level training.