CUA-Gym用三个AI智能体自动生成训练数据，省去人工标注。它克隆了94个常用软件，训练出的模型追平了Claude Sonnet 4.6，小模型3B参数达到17B效果，还完全开源。
CUA-Gym是一个端到端流水线，通过三个协调编码智能体自动生成可验证的计算机使用任务。它构建了94个流行应用的模拟版本，包括Slack、Notion、Salesforce和Gmail克隆，并直接读取状态设计奖励函数。生成的训练数据集包含32,112个验证元组，覆盖110个环境。基于该数据训练的模型在OSWorld-Verified基准上达到72.6%，与Claude Sonnet 4.6持平。一个3B参数的小模型以十分之一参数匹配了17B基座模型的性能。
The biggest bottleneck for computer-use agents jus…
The biggest bottleneck for computer-use agents just got automated away.
Reinforcement learning broke open math and coding.
But for agents clicking around real software, progress stalled. The bottleneck was generating training data at scale. CUA-Gym is a pipeline that solves this.
It synthesizes verifiable tasks for computer-use agents end to end. The setup uses three coordinated coding agents: > Generator writes environment setup scripts > Discriminator drafts the reward function blind > Orchestrator iterates until both align The team also built mock versions of 94 popular apps.
These include Slack, Notion, Salesforce, and Gmail clones. Rewards read state directly, skipping flaky screenshot judges. The resulting dataset holds 32,112 verified tuples across 110 environments. A trained model hits 72.6% on OSWorld-Verified, matching Claude Sonnet 4.6.
A smaller 3B version matches its 17B base with 10x fewer parameters. The full system, dataset, and models are open source.