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

A new training method called Category-Aware Iterative Expert Training (One to More, More to One) targets uneven progress in repository-level software engineering agents, where pooled agentic reinforcement learning produces gains in some task categories alongside regressions in others. The approach addresses how aggregate resolution rates obscure these category-level changes in SWE 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 s