Training Software Engineering Agents and Verifiers with SWE-Gym
Jiayi Pan*, Xingyao Wang*, Graham Neubig, Navdeep Jaitly, Ji Heng, Alane Suhr^, Yizhe Zhang^
UC Berkeley, UIUC, CMU, Apple | *, ^ denotes equal contribution
We present SWE-Gym, the first environment for training real-world software engineering agents. We use it to train strong LM agents that achieve state-of-the-art open results on SWE-Bench, with early, promising scaling characteristics as we increase training and inference-time compute.