‹ back to 2026-09-21

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

arXiv cs.CL LLM Agents Shuai Bai, Jiayong Deng, Yikun Fu, Chang Gao, Xuhao Hu, Mianqiu Huang, Yizhen Jiang, Yuheng Jing, Dehui Kong, Keliang Li, Ning Li, Wanli Li, Dayiheng Liu, Dunjie Lu, Changwei Luo, Que Shen, Zheyuan Wang, Zijian Wang, Jie Wu, Gao Wu, Zhihui Xie, Rui Xie, Haiyang Xu, An Yang, Jiakang Yuan, Yanming Zhang, Jiajun Zhang, Xi Zhang, Zhenru Zhang, Zhuo Zhen, Mingkang Zhu, Bowen Zhou 2026-09-18

TL;DR - RecreationWorld is a cross-platform framework for training and evaluating hybrid computer-use agents that combine GUI exploration, coding, execution, and visual verification. Its results show meaningful transfer from recreation-based training, but also reveal substantial gaps in reproducing interactive behavior and computed outputs.

RecreationWorld supports reproducible tasks across Ubuntu, macOS, Windows, Android, and Web through unified GUI-control and coding tools.
Training on trajectories generated from open-source applications improved performance across five out-of-distribution coding and hybrid computer-use benchmarks.
RecreationBench contains 250 held-out tasks with reference-validated programmatic and visual assertions spanning multiple interaction depths.
GPT-6 Astra scored 58.1% overall but passed every programmatic test on only 2.8% of tasks; agents handled static interfaces better than interactions and computed outputs.

view merged work →