RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents
RecreationBench scores hybrid agents on whether they can recreate working software from a live reference, not just copy its surface.
The paper introduces RecreationWorld, a five-platform framework spanning Ubuntu, macOS, Windows, Android, and Web. Agents inspect a running reference app, choose when to use GUI actions or code, then build and visually verify their own implementation. Its held-out benchmark has 250 tasks with programmatic and visual tests validated against the reference and by human reviewers. GPT-6 Astra leads with 58.1% overall, but passes all programmatic tests on only 2.8% of tasks. HF Daily Papers' note
The paper introduces RecreationWorld, a five-platform framework spanning Ubuntu, macOS, Windows, Android, and Web. Agents inspect a running reference app, choose when to use GUI actions or code, then build and visually verify their own implementation. Its held-out benchmark has 250 tasks with programmatic and visual tests validated against the reference and by human reviewers. GPT-6 Astra leads with 58.1% overall, but passes all programmatic tests on only 2.8% of tasks. HF Daily Papers' note
score 5