A task description alone is insufficient for training a computer-use agent: the environment must reset correctly and success must be checkable. The key idea is to generate the task, world state, and reward together while isolating the reward writer from the setup code.
Key Findings
- Independent generator and discriminator agents reduce rewards that merely mirror the construction procedure; rollout filtering checks that tasks remain solvable.
- The dataset contains 32,112 verified tuples across 16 desktop and 94 mock-web environments, broadening training beyond a few applications.
- Using 10,858 tuples for RL, the A3B and A17B models score 62.1% and 72.6% on OSWorld-Verified; held-out WebArena gains suggest transfer beyond the mocks.
- Environment diversity adds value beyond more trajectories, making coverage a separate scaling lever.
Who Should Read It
Great fit if you build GUI-agent training data or study verifiable rewards. Look elsewhere if you need deployment guarantees: endpoint rewards can miss destructive intermediate actions, mocks omit real authentication and service failures, and the largest RL runs use single seeds.