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

2026-06-22

AI Summary

This work addresses the challenge faced by multimodal GUI agents in real-world desktop environments, where sparse, delayed, and costly feedback hinders the discovery of successful trajectories. To overcome this, the authors propose the Environment-Native Validation Search (ENVS) framework, which introduces an in-environment validation mechanism for the first time. Within the OSWorld virtual machine, ENVS enables efficient policy optimization through action-branch exploration, validation of successful paths, and the construction of step-level supervision signals. The study also introduces OSWorld-Noisy, a dynamic perturbation benchmark designed to evaluate agent robustness. Experimental results demonstrate that ENVS achieves a pass@8 rate of 30.3 across 300 tasks (29.0 under perturbations), outperforming online reinforcement learning methods such as ARPO, significantly reducing GPU overhead, and excelling in vision-assisted reasoning tasks.

Abstract

As multimodal agents move from interface understanding to real software control, successful trajectory discovery in live desktop environments becomes a key challenge. GUI tasks require long-horizon sequences of precise mouse and keyboard actions, while feedback is sparse, delayed, and costly to obtain through VM rollouts. We propose Environment-Native Verified Search (ENVS), a training-time search-and-filter pipeline that uses the environment to construct verified supervision before policy optimization: it branches over behaviorally distinct GUI actions in live OSWorld VMs, verifies successful leaves, and trains from globally balanced step-level supervision. To evaluate robustness under realistic desktop interruptions, we also introduce OSWorld-Noisy, a dynamic benchmark for recoverable desktop interruptions that preserves the original tasks while testing whether agents can refocus, dismiss, wait, or recover under live perturbations. On the 300-task OSWorld pool, ENVS reaches 30.3 pass@8 on original evaluations and 29.0 on OSWorld-Noisy, outperforming matched ARPO-style online RL while reducing compute from 184-192 to 138-153 GPU-hours; even with only 30% of its search data, ENVS reaches 27.0 pass@8, exceeding ARPO from the base model. Training from noisy environments also better preserves visual-reasoning abilities on auxiliary benchmarks, including OSWorld-G Refusal (16.7 vs. 1.9) and BLINK Functional Correspondence (26.2 vs. 23.1).

Problem

• long-horizon GUI tasks
• sparse feedback
• desktop environment
• trajectory discovery
• multimodal agents

Innovation

• Environment-Native Verified Search
• long-horizon GUI agents
• verified supervision
• OSWorld-Noisy
• step-level supervision