Paper

Agent World Model in the Wild: Computer-Use Agent Environment Synthesis

Zhaoyang Wang, Qianhui Wu, Haorui Wang, Simon Yu, Baolin Peng, Canwen Xu, Hao Cheng, Wenlin Yao

Builds open-source synthetic PC worlds with GUI, CLI, and MCP access, then generates and validates cross-interface tasks via multi-agent execution gates.

arXiv ↗Oct 19, 2026Code ↗

llm
agents
reinforcement learning
environment generation

Full pipeline: world synthesis, XAI task synthesis, gates, online RL for CUA

What it is

Synthesizes executable PC worlds and cross-interface tasks for training computer-use agents, using only open-source models and execution-based quality gates. Unlike prior single-interface or proprietary pipelines, it builds unified PC environments spanning GUI, MCP, and CLI, then filters synthesized long-horizon tasks with execution-based gates for solvability, difficulty, and verifier reliability.

Key results

341 long-horizon environment-task pairs synthesized
Tasks can require 100+ steps; GPT-6-Astra solves only 60%
Qwen models trained only in WildAWM improve on 8/8 OOD benchmarks

Open-source synthetic computer worlds can train agents that generalize broadly beyond the generated environments.

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

What it is

Key results

One useful ML email a week.

More from Oct 19 – Oct 25

Worth your time, not all of your time.