WildAWM builds 341 agent-training worlds from open-source models

WildAWM, a pipeline described in a new arxiv paper, synthesizes 341 long-horizon computer-use training environments using only open-source models, with tasks hard enough that GPT-6-Astra, per the abstract, is reported as "solving only 60% of them."

The motivation is written plainly in the opening: "scaling such learning directly in real-world computer systems is challenging and can introduce substantial security risks." Rather than let agents loose on live machines, the authors (a group including Zhaoyang Wang, Qianhui Wu, Jianfeng Gao, and Huaxiu Yao) construct executable environments that "jointly support graphical user interfaces, model context protocol services, and command-line interfaces," letting trained agents combine visual interaction, tool use, and code execution.

The harder problem is quality control at scale. The paper lists three recurring task-synthesis challenges (solvability, sufficient difficulty, and reliable verifiers) and tackles them with "execution-based quality gates that combine heterogeneous agent rollouts, cross-judge consistency checks, and counterfactual probes of judgment robustness." Some tasks are described as requiring over 100 agent steps.

Two Qwen models of different sizes, trained "exclusively in these synthetic worlds," are reported to make "substantial gains across 8 out-of-distribution benchmarks spanning web, claw, and computer-use agents." Code is published at github.com/MSR-Orchard/WildAWM.