GUI-Owl-1.5 稱霸 20 多項 GUI 基準

Post LinkedIn

📄閱讀原文： ArXiv AI

#gui-agents#multi-platform#rl-scalinggui-owl-1.5

💡Open-source GUI agent hits SOTA on 20+ benchmarks across desktop/mobile/web (71.6 AndroidWorld)

⚡ 30-Second TL;DR

有什麼變化

SOTA 分數：OSWorld 56.5、AndroidWorld 71.6、WebArena 48.4

為什麼重要

推進跨平台 GUI 自動化，提升即時代理應用。開源賦能開發者微調 SOTA 模型用於自訂案例。

下一步行動

Clone https://github.com/X-PLUG/MobileAgent and test the online cloud-sandbox demo.

誰應關注：Researchers & Academics

關鍵要點

•SOTA 分數：OSWorld 56.5、AndroidWorld 71.6、WebArena 48.4
•指令/思考變體，尺寸 2B/4B/8B/32B/235B
•混合數據飛輪，利用模擬/沙盒環境收集數據
•MRPO RL 演算法處理多平台長視野任務
•開源模型及 GitHub 雲沙盒演示

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 5 個來源。

🔑 增強重點摘要

•GUI-Owl-1.5 achieves state-of-the-art performance among open-source models on over 20 GUI benchmarks, including 56.5 on OSWorld, 71.6 on AndroidWorld, and 48.4 on WebArena[1].
•The 8B-Thinking variant excels with 52.9 on OSWorld-Verified, surpassing UI-TARS-2 (53.1) at similar scale and general-purpose models like Qwen3-VL-235B-A22B-Think (38.1)[1].
•On browser benchmarks, GUI-Owl-1.5-8B-Thinking scores 46.7 on WebArena, 40.8 on VisualWebArena, 78.1 on WebVoyager, and 48.6 on Online-Mind2Web, outperforming other open-source models[1].
•Thinking variants consistently outperform Instruct versions, especially on long-horizon planning tasks like WebVoyager (69.9 to 82.1) and Online-Mind2Web (41.7 to 48.6)[1].
•GUI-Owl models (e.g., 7B and 32B) show strong prior results on GUI grounding benchmarks like ScreenSpot-Pro and others, with scores up to 99.0 and 96.1 across tasks[2].

📊 競品分析▸ Show

Model	Key Benchmarks	Scale	Notes
GUI-Owl-1.5-8B-Thinking	OSWorld-Verified: 52.9, WebArena: 46.7, WebVoyager: 78.1, Online-Mind2Web: 48.6	8B	SOTA open-source multi-platform GUI agent[1]
UI-TARS-2	OSWorld-Verified: 53.1	Comparable to 8B	Slightly higher on OSWorld but narrower scope[1]
Qwen3-VL-235B-A22B-Think	OSWorld-Verified: 38.1	235B	General-purpose, lags on GUI tasks[1]
GUI-Owl-7B	ScreenSpot-Pro variants: 64.8-86.4, grounding: up to 99.0	7B	Strong prior grounding performance[2]
UI-Venus-7B	ScreenSpot-Pro: 74.6	7B	Competitive but outperformed by GUI-Owl-32B[2]

🛠️ 技術深入

No specific details on model architecture, Hybrid Data Flywheel, or MRPO RL algorithm found in search results. Benchmarks confirm multi-platform (desktop, mobile, browser) evaluation across automation, grounding, tool calling, memory, and knowledge tasks[1].

🔮 前景展望AI analysis grounded in cited sources

GUI-Owl-1.5 advances open-source GUI agents for cloud-edge collaboration, enabling stronger automation on OSWorld, AndroidWorld, and WebArena, potentially accelerating multi-platform agent deployment while competing with proprietary systems[1].