Researchers at Shanghai Jiao Tong University published SoftwareGym, an automated pipeline for creating software environments to train computer-use agents, on October 8. The paper says building native application environments requires substantial construction and integration work, limiting coverage and increasing the risk that agents overfit to specific applications. Browser-hosted environments, meanwhile, offer limited interaction depth and workflow complexity.

SoftwareGym generates structurally diverse designs, turns them into runnable applications with varied interfaces and executable functions, then tests their behavior through GUI interactions and iterative repairs. The researchers synthesized 182 applications across twelve categories, reporting a mean GUI task pass rate of 91.5%. They also created SoftwareGym-Bench, a 120-task evaluation suite spanning the same twelve categories, and collected interaction trajectories from those environments for agent training.

A model trained with the collected trajectories improved Pass@3 on the strictly out-of-domain OSWorld benchmark from 41.39% to 56.97% and scored 40.0% on SoftwareGym-Bench. The authors say the results show synthesized software environments can provide scalable, transferable supervision for computer-use-agent training.

Researchers at Shanghai Jiao Tong University published SoftwareGym, an automated pipeline for creating software environments to train computer-use agents, on October 8. The paper says building native application environments requires substantial construction and integration work, limiting coverage and increasing the risk that agents overfit to specific applications. Browser-hosted environments, meanwhile, offer limited interaction depth and workflow complexity.

SoftwareGym generates structurally diverse designs, turns them into runnable applications with varied interfaces and executable functions, then tests their behavior through GUI interactions and iterative repairs. The researchers synthesized 182 applications across twelve categories, reporting a mean GUI task pass rate of 91.5%. They also created SoftwareGym-Bench, a 120-task evaluation suite spanning the same twelve categories, and collected interaction trajectories from those environments for agent training.

A model trained with the collected trajectories improved Pass@3 on the strictly out-of-domain OSWorld benchmark from 41.39% to 56.97% and scored 40.0% on SoftwareGym-Bench. The authors say the results show synthesized software environments can provide scalable, transferable supervision for computer-use-agent training.