This paper teaches code AIs to work more like real software engineers by training them in the middle of their learning using real development workflows.
SAGE is a two-agent system that automatically writes tough, multi-step search questions and checks them by actually trying to solve them.
The paper introduces UCoder, a way to teach a code-generating AI to get better without using any outside datasets, not even unlabeled code.