WildAWM
An open-source synthetic environment generation framework proposed by Microsoft Research that uses open models to automatically generate executable computer-agent training environments supporting three interaction modes—GUI, MCP and CLI—together with an execution-based quality gating mechanism for filtering tasks.
Timeline (last 90 days)
Microsoft Research proposed WildAWM (Agent World Model in the Wild), a synthetic pipeline that uses open-source models to automatically generate controllable computer environments and tasks for training computer agents.
The WildAWM code has been open-sourced on GitHub, with the repository path MSR-Orchard/WildAWM.
The executable environments built by WildAWM support three interaction modes simultaneously: graphical user interface (GUI), Model Context Protocol (MCP) service, and command-line interface (CLI).
WildAWM introduces execution-based quality gating, comprising three mechanisms: heterogeneous agent rollout, cross-judge consistency checking, and counterfactual probes.
WildAWM produced 341 environment-task pairs.
The tasks in WildAWM are of the long-horizon type, and some tasks require more than 100 agent steps to complete.
Even the frontier model GPT-6-Astra can solve only 60% of the WildAWM tasks.
The research team trained two Qwen models of different sizes solely on the WildAWM synthetic world and evaluated them on 8 out-of-distribution (OOD) benchmarks, covering web agents, claw-type tasks, and computer-use agents.
Both Qwen models trained solely in WildAWM synthetic environments achieved significant improvements on all 8 OOD benchmarks, indicating that the acquired capabilities generalize to unseen computer-use tasks.
The paper lists how to quantify the distribution gap between synthetic environments and real systems as one of the open problems.
2 more timeline events
All Facts (12)
The paper lists whether 341 task pairs are sufficient to support larger-scale training as one of the open problems.
50%UnverifiedMicrosoft Research proposed WildAWM (Agent World Model in the Wild), a synthetic pipeline that uses open-source models to automatically generate controllable computer environments and tasks for training computer agents.
50%UnverifiedThe executable environments built by WildAWM support three interaction modes simultaneously: graphical user interface (GUI), Model Context Protocol (MCP) service, and command-line interface (CLI).
50%UnverifiedWildAWM introduces execution-based quality gating, comprising three mechanisms: heterogeneous agent rollout, cross-judge consistency checking, and counterfactual probes.
50%UnverifiedWildAWM produced 341 environment-task pairs.
50%UnverifiedThe tasks in WildAWM are of the long-horizon type, and some tasks require more than 100 agent steps to complete.
50%UnverifiedEven the frontier model GPT-6-Astra can solve only 60% of the WildAWM tasks.
50%UnverifiedThe research team trained two Qwen models of different sizes solely on the WildAWM synthetic world and evaluated them on 8 out-of-distribution (OOD) benchmarks, covering web agents, claw-type tasks, and computer-use agents.
50%UnverifiedBoth Qwen models trained solely in WildAWM synthetic environments achieved significant improvements on all 8 OOD benchmarks, indicating that the acquired capabilities generalize to unseen computer-use tasks.
50%UnverifiedThe paper lists how to quantify the distribution gap between synthetic environments and real systems as one of the open problems.
50%UnverifiedThe paper lists how to maintain the reliability of verifiers on more complex tasks as one of the open problems.
50%UnverifiedThe WildAWM code has been open-sourced on GitHub, with the repository path MSR-Orchard/WildAWM.
50%