ENVS: Environment-Native Verified Search for Long-Horizon GUI Agents

The article introduces Environment-Native Verified Search (ENVS), a novel approach for training long-horizon GUI agents that leverages environment feedback to enhance supervision during policy optimization. ENVS utilizes a search-and-filter pipeline in live OSWorld VMs to verify successful GUI actions, achieving a pass rate of 30.3 on the 300-task OSWorld benchmark and 29.0 on the OSWorld-Noisy benchmark, while significantly reducing computational costs from 184-192 GPU hours to 138-153. This method not only improves task performance but also enhances visual-reasoning capabilities in noisy environments, making it a valuable advancement for practitioners developing robust AI agents for real-world applications.