WildAWM synthesizes environments for computer-use agent training
Original title: Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
The open-source pipeline made 341 long-horizon tasks; GPT-6-Astra solved 60%, and two trained Qwen models improved on eight benchmarks.
WildAWM is a pipeline that synthesizes controllable environments and tasks for training computer-use agents without running that learning on live systems. Environments jointly support graphical interfaces, model context protocol services, and command lines, and generation uses only open-source models. Execution-based quality gates produced 341 long-horizon pairs that can take over 100 steps; the authors report GPT-6-Astra solved 60%. Two Qwen models trained only in these worlds improved on 8 out-of-distribution web, claw, and computer-use benchmarks.
Unconfirmed: The 60% solve rate is the paper's reported result for GPT-6-Astra on the synthetic tasks.
- Quality gates combine heterogeneous rollouts, cross-judge checks, and counterfactual probes.
- Tasks stay difficult for frontier models even after synthesis filters.
- Both Qwen sizes gained on web, claw, and computer-use benchmarks.
- Code is public at github.com/MSR-Orchard/WildAWM.
- Products
- WildAWM
- AI models
- GPT-6-AstraQwen
From the original article
1 min read at arxiv.orgAgent learning is shifting from learning from human data toward learning through interaction with the world. However, scaling such learning directly in real-world computer systems is challenging and can introduce substantial security risks.
By Zhaoyang Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.