SoftwareGym：用合成软件环境扩展计算机使用智能体训练
原标题：SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments
AI 摘要
研究者提出 SoftwareGym，一个自动化多智能体流水线，用于合成具有多样界面、可执行功能和运行时验证行为的 GUI 软件环境，以解决计算机使用智能体（CUA）训练中环境构建成本高、覆盖有限的问题。该流水线合成了横跨 12 个类别的 182 个可运行应用，平均 GUI 任务通过率达 91.5%，并构建了 120 任务的 SoftwareGym-Bench 评测集。用这些环境收集的轨迹训练模型后，在严格域外的 OSWorld 基准上 Pass@3 从 41.39% 提升至 56.97%，在 SoftwareGym-Bench 上达到 40.0%。
正文节选
SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments Abstract Computer-use agents (CUAs) are rapidly improving, yet synthesizing diverse and reliable software environments at scale remains challenging. Native application environments require substantial construction and integration effort, which limits application coverage and increases the risk of application-specific overfitting. Browser-hosted environments provide limited interaction depth and workfl