RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents
RecreationWorld offers verifiable five-platform environments for hybrid computer-use agents; GPT-6 Astra tops RecreationBench at 58.1% overall.
RecreationWorld is a five-platform framework (Ubuntu, macOS, Windows, Android, Web) where hybrid agents must discover a running reference's behavior and rebuild it, with the reference serving as an oracle for hidden behavioral tests providing execution-grounded rewards. Models trained on scaled trajectories improve across five out-of-distribution coding and hybrid computer-use benchmarks and verify rendered outputs more often. RecreationBench adds 250 held-out tasks with reference-validated programmatic and visual assertions; GPT-6 Astra leads at 58.1% overall but passes all programmatic tests on only 2.8% of tasks.
- Hybrid agents blend GUI interaction, coding, and visual verification
- Five platforms with reference application as behavioral oracle
- RecreationBench: 250 tasks, GPT-6 Astra leads at 58.1%
- Agents reproduce static UI structure better than interactions and computed outputs
From the original article
1 min read at huggingface.co
Computer-use agents (CUAs) have advanced along two separate lines: graphical interaction and software development through code and the command line. Real digital work requires both, interleaved rather than stacked end to end. We study hybrid CUAs that autonomously decide when to explore an interface, implement software, and run and visually verify their artifacts.
By Shuai Bai, Jiayong Deng, Yikun Fu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.