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 frame
Reading in Plain English mode — toggle below to switch
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 frame
This work addresses a critical challenge in training AI agents for complex software engineering tasks: managing performance across diverse task categories. By introducing category-aware expert training and policy integration, it offers a method to prevent performance regressions in some areas while improving others, leading to more robust and reliable agents.