Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
Two Qwen models trained exclusively in synthetic computer environments improved across eight benchmarks of unseen tasks, the authors report.
WildAWM uses open-source models to build executable environments supporting graphical interfaces, model context protocol services and command-line interaction. Its 341 environment-task pairs can require more than 100 agent steps; GPT-6-Astra solved only 60% of them. Execution-based checks test task solvability, difficulty and verifier reliability. The pipeline is designed to support agent training without the security risks of interaction with real computer systems.
HF Daily Papers' note
WildAWM uses open-source models to build executable environments supporting graphical interfaces, model context protocol services and command-line interaction. Its 341 environment-task pairs can require more than 100 agent steps; GPT-6-Astra solved only 60% of them. Execution-based checks test task solvability, difficulty and verifier reliability. The pipeline is designed to support agent training without the security risks of interaction with real computer systems.
HF Daily Papers' note
score 5