SoftwareGym: synthetic apps to train computer-use agents lift an out-of-domain score

Researchers at Shanghai Jiao Tong University introduce SoftwareGym, featured on Hugging Face Daily Papers. Training AI agents to use software needs many realistic apps to practise in, which are costly to build. SoftwareGym instead generates them automatically: it designs varied interfaces, turns them into working applications, and checks each one works by running GUI tasks and repairing failures. The team made 182 apps across twelve categories, with a 91.5% task pass rate. Using data collected from these synthetic apps to train a computer-use model, they report its Pass@3 score on OSWorld, a separate real-world benchmark the apps were not drawn from, rose from 41.39% to 56.97%. This is a research result; the authors link code on GitHub.