WildAWM Builds Executable Worlds for General Computer-Use Agents

Introduction

General computer-use agents need more than the ability to click buttons or invoke an API. They must maintain state, choose among interfaces, recover from errors, and complete goals that unfold over many steps. Collecting this experience directly on real computer systems is expensive and can expose private data or create security risks. WildAWM, introduced by a Microsoft Research team, explores a different route: training agents through interaction with executable synthetic worlds.

What WildAWM contributes

• One world, multiple interfaces. The generated environments are persona-driven personal computers that can evolve over multiple days. Websites, MCP services, and local files share a common state. An agent can therefore combine visual interaction, tool calls, and command-line execution within the same task rather than being locked into a single benchmark interface.
• Open-source generation. The pipeline is powered entirely by open-source models. This is important for reproducibility and cost control: large collections of interactive training scenarios do not have to depend on proprietary frontier models.
• Execution-based quality control. Synthetic tasks often fail in three ways: they may be impossible, too easy, or judged incorrectly. WildAWM addresses these problems with actual rollouts from heterogeneous agents, consistency checks across judges, and counterfactual probes that test whether a verifier remains reliable when relevant conditions are changed.
• Long-horizon difficulty. The released collection contains 341 environment-task pairs. Some require more than 100 agent steps, while the reported success rate for GPT-6-Astra is about 60 percent. The benchmark is therefore intended to test planning and state management, not just interface familiarity.
• Transfer beyond the synthetic world. Two Qwen models of different sizes were trained exclusively in WildAWM environments. Across eight out-of-distribution benchmarks covering web, claw, and computer-use agents, Qwen3.6-35B-A3B and Qwen3.5-4B improved by 10.58 and 8.21 points on average, respectively.

Why it matters

The central contribution is a closed loop connecting world generation, task synthesis, execution, and verification. For computer-use agents, that loop can provide a safer place to fail, replay trajectories, and adjust difficulty than real applications. It also offers a way to generate experiences where file state, tools, websites, and commands affect one another—conditions that are closer to practical automation than isolated web tasks.

The results point to a shift in the data bottleneck. More model parameters alone may not solve the problem if training tasks teach only narrow interface habits. Shared state and long sequences force an agent to plan, remember, select tools, and recover from intermediate mistakes. At the same time, synthetic worlds can still miss the unpredictability and breadth of real software environments. Their long-term value will depend on how well generated scenarios cover those gaps and how consistently the reported gains hold under independent evaluation. WildAWM is best understood as a scalable training infrastructure, not a complete solution to general computer use.

Source: Hugging Face Daily Papers