RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents

Computer-use agents increasingly need to combine graphical interaction with software development through code and command-line tools. RecreationWorld studies this hybrid setting through a five-platform framework built around recreation: given a running reference application, an agent must explore its behavior, implement a faithful version, and verify the result with native GUI control and coding tools. The framework supports Ubuntu, macOS, Windows, Android, and Web environments, and introduces RecreationBench, a held-out benchmark of 250 tasks with reference-grounded programmatic and visual assertions. Results show that even strong frontier models remain far from reliable recreation, highlighting open challenges in long-horizon planning, multimodal grounding, interaction fidelity, and execution-grounded verification.

arXiv preprint