News · 2026-10-22

WildAWM trains computer agents inside replayable synthetic workplaces

WildAWM generates executable synthetic computer environments and reports that agents trained in them improve on eight external benchmarks. Its October 19 paper describes 341 accepted tasks with persistent files, service data and websites, rather than a screen predictor alone. The gains are bounded: the authors find no clear improvement on the hardest task quartile, where the teacher also struggles.

Key facts

• Zhaoyang Wang and coauthors posted the paper on October 19, 2026.
• The pipeline generated 640 world states and retained 341 environment–task pairs.
• The reported cross-benchmark average improvement for the larger model is 10.58 points, mixing different benchmark score types.
• The primary source is the WildAWM paper.

A computer agent needs practice that is both challenging and checkable. Real offices contain private information, brittle services and actions that cannot be reset casually. A synthetic workplace can be cheaper to repeat, but only if the tasks correspond to functioning software and success means something. A convincing screenshot is insufficient when the database never records the action the screen claims succeeded.

Wang and colleagues call their work “Computer-Use Agent Environment Synthesis” in the paper’s title. The full text describes expanding personas into sequences of day plans. A coding agent materializes those plans as deterministic changes to files, databases and mirrored websites. The resulting environment can accumulate history rather than presenting every task as an unrelated empty desktop.

Think of a rehearsal office built for training new staff. It contains working forms, appointment records, documents and service counters. A trainee must actually change the relevant records, and the instructor can reset the office afterward. This differs from watching a film of an office and guessing the next frame. WildAWM’s environments have executable state, so agents can make decisions whose consequences can be checked.

The pipeline uses 36 personas and reports 640 world states, but those are not 640 accepted training tasks. The authors retain a state only after independent replays agree, services remain healthy and another coding agent checks consistency with the plan. The final task set contains 341 accepted pairs. Keeping these counts separate prevents synthesis volume from being mistaken for usable training volume.

Tasks span graphical interfaces, service tools and command lines. Their average includes 3.1 interface switches, and many return to an interface already visited. This matters because a real workflow may require taking information from a website, editing a local artifact and updating a service without losing track of what happened. The benchmark is attempting to train those dependencies, not merely isolated clicks.

Each task has an executable checker comparing relevant initial and final state and the agent’s answer. All binary criteria must pass for full task success, while partial checklist scores can support reinforcement learning. The pipeline checks that a task is initially unsatisfied, solvable and sufficiently difficult, and examines agreement between checkers and model-assisted judges. It also probes whether a judge maintains its verdict when challenged with an opposing claim.

The paper documents why these gates matter. A page can announce success without persisting a database row. A checker can accept a stale entry. A supposedly essential interface can turn out to be avoidable because another message repeats its information. Each error would distort the lesson learned or the score assigned. This connects directly to reward hacking: an agent improves what the grader rewards, including flaws in the grader.

The authors trained two model sizes on the accepted tasks, using separate portions for demonstration learning, reinforcement learning and internal evaluation. Their reported external improvements average 8.21 points for the smaller model and 10.58 for the larger one. Those averages combine pass rates with other mean-score metrics. They should be read as the paper’s cross-benchmark summary, not a uniform increase in accuracy across identical tests.

The strongest limitation comes from the authors’ difficulty analysis. Improvements concentrate where teacher demonstrations provide useful guidance, with no clear gain in the hardest quartile. Some comparison rows also come from earlier papers rather than controlled reruns in matching software. These details narrow the result to transfer under the tested training and evaluation setup, rather than evidence that agents independently acquired arbitrary new computer skills.

The released repository documents the container-based runtime and model endpoint configuration. It is research infrastructure, not evidence of a turnkey workplace assistant. The dossier found early attention but no independent replication. WildAWM’s contribution is a way to manufacture more inspectable practice environments; its next test is whether that practice transfers reliably beyond the demonstrated workflows. The existing lesson on synthetic data explains why generation quality and validation must scale together.

Primary source, verified: read the paper → (arXiv 2610.26828)

Ground Truth. (2026, October 22). WildAWM trains computer agents inside replayable synthetic workplaces. Ground Truth. https://groundtruth.day/news/wildawm-executable-agent-training-worlds.html

@misc{groundtruth:wildawm-executable-agent-training-worlds,
  title  = {WildAWM trains computer agents inside replayable synthetic workplaces},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/wildawm-executable-agent-training-worlds.html}
}

Topics: agents · synthetic-data · computer-use · evaluation · world-models