Long-horizon GUI automation fails in practice because training data are scarce, biased, and often omit tacit user conventions; that makes execution brittle even for strong base models. The paper's core insight is that coupling a scalable environment-grounded training pipeline with in-context, subtask-level demonstrations yields much higher reliability: demonstration guidance raises strict success and task progress substantially on long multi-step office tasks.
Key Findings
- Strong empirical gains: the 27B agent variant achieves 77.0% on OSWorld-Verified and 66.2% on WindowsAgentArena, and on the new OSWorkerBench (100 long-horizon office tasks) it reaches 41.0% strict success and 76.9% progress—material improvements over its Qwen3.6-27B base (±17.7 and ±24.5 points).
- Demonstration impact: providing one successful, closely related demonstration on a 33-task self-demo subset improves strict success from 17.2% to 35.4% and progress from 67.9% to 81.1%, showing in-context demos help bridge unstated conventions and tool-specific details.
- System design: a closed-loop data engine automates task generation, environment construction, rollout, filtering, capability balancing, SFT and online RL across parallel environments; an in-context demonstration mechanism converts multimodal traces into reusable, subtask-level workflows and supports live replanning.
Who it's for & Tradeoffs
Great fit if you research or build multimodal GUI agents, need a reproducible benchmark for long-horizon desktop automation, or want to evaluate how demonstrations improve open-weight foundation agents. Look elsewhere if you need a lightweight on-device solution or prefer methods that avoid large synthetic rollout infrastructures. Expect substantial compute and engineering overhead: parallel environment orchestration, verifier tooling, and large-model SFT/RL training are central to the approach.
Where It Fits
Compared to prior GUI-agent work, this paper emphasizes a foundation-model stack + in-context demonstrations and a large, focused benchmark. It complements works that target memory-based experience reuse, few-shot demonstration parsing, or vision-language-action architectures by integrating scalable environment grounding, demonstration-guided execution, and an evaluation suite for long workflows.