Agent World Model in the Wild, or WildAWM, proposes generating controlled environments for training computer-use agents. The motivation is to give agents room to learn through interaction without doing all that experimentation directly in real computer systems.
More than one interaction interface
The generated environments support graphical interfaces, Model Context Protocol services and command-line interaction together. An agent can therefore combine visual actions, tool calls and code execution within the same environment.
The authors say the pipeline uses open-source models to generate these worlds. That claim concerns the environment-generation process; the paper also evaluates the resulting tasks with a frontier model.
How the tasks are checked
WildAWM addresses three problems in synthetic task generation: whether a task can be solved, whether it is sufficiently difficult and whether its verifier judges the result reliably. Its execution-based checks combine different agent rollouts, consistency checks between judges and counterfactual probes of judging robustness.
The resulting collection contains 341 environment-task pairs, with some requiring more than 100 agent steps. The paper reports that GPT-6-Astra solved 60% of them. That is a result on this generated collection, rather than a general computer-use success rate.
What the training result suggests
The researchers trained two Qwen models of different sizes exclusively in the synthetic worlds and report gains across eight out-of-distribution benchmarks. This supports examining synthetic interaction as a training resource, while leaving the details of transfer important.
Franklin’s interpretation is that verifier quality deserves as much scrutiny as environment variety. If a task can be completed through an unintended shortcut, its training signal may reward behaviour that does not transfer to the intended workflow.
For a practical assessment, inspect the allowed interfaces, success conditions and failure cases alongside benchmark scores. Results in controlled worlds do not establish safe operation in a live organisation with real accounts and irreversible actions.
Read WildAWM for the pipeline and reported evaluation.
Comments