WildAWM synthesizes computer-use agent environments with open models
Original title: Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
It yields 341 tasks that can exceed 100 steps; GPT-6-Astra solves 60%, and Qwen models trained only there improve on eight benchmarks.
Researchers propose Agent World Model in the Wild (WildAWM), a pipeline that synthesizes controllable environments and tasks for computer-use agents. Environments jointly support graphical interfaces, model context protocol services, and command-line interfaces, using only open-source models. Execution-based quality gates check solvability, difficulty, and verifier reliability. Two Qwen models trained only in these worlds improved on eight out-of-distribution benchmarks spanning web, claw, and computer-use agents.
Unconfirmed: Benchmark scores are the authors' reported results and are not independently confirmed.
- The pipeline produced 341 environment-task pairs that can exceed 100 agent steps.
- GPT-6-Astra solved only 60% of the synthesized tasks.
- Quality gates combine heterogeneous rollouts, cross-judge checks, and counterfactual probes.
- Public code is at github.com/MSR-Orchard/WildAWM.
- Organizations
- MSR-Orchard
- AI models
- QwenGPT-6-Astra
From the original article
1 min read at huggingface.co
Agent 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, Qianhui Wu, Haorui Wang · huggingface.co
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.