Starting from generated user personas and daily activities, WildAWM constructs executable personal-computer environments containing websites, local files and model context protocol services.
Synthetic computer worlds improve agents on unseen computer-use tasks
WildAWM generates executable, cross-interface training environments whose tasks improved two Qwen models across eight out-of-distribution benchmarks.
Big Tech
Zhaoyang Wang · Qianhui Wu · Haorui Wang · Simon Yu · Baolin Peng · Canwen Xu · +6 more
UNC-Chapel Hill · Microsoft Research · Georgia Tech · Northeastern University · Independent Researcher
Research Digest··3 min read
The authors developed WildAWM, an open-source pipeline that synthesizes computer environments, tasks and automated success checks for training computer-use agents.
§