SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments
SoftwareGym reports 182 synthesized GUI apps and a 91.5% mean task pass rate for generated environments.
The paper frames synthesized software as a way around costly native app setup and shallow browser-only training tasks. Its pipeline generates runnable GUI applications, checks behavior at runtime, and repairs failures through GUI evaluation. The authors also introduce SoftwareGym-Bench, a 120-task suite across twelve categories. Training on collected trajectories raised Pass@3 on OSWorld from 41.39% to 56.97%, according to the paper. HF Daily Papers' note
The paper frames synthesized software as a way around costly native app setup and shallow browser-only training tasks. Its pipeline generates runnable GUI applications, checks behavior at runtime, and repairs failures through GUI evaluation. The authors also introduce SoftwareGym-Bench, a 120-task suite across twelve categories. Training on collected trajectories raised Pass@3 on OSWorld from 41.39% to 56.97%, according to the paper. HF Daily Papers' note
score 5