The paper introduces GUI-Owl-1.5, a native GUI agent model featuring instruct and thinking variants in sizes ranging from 2B to 235B parameters. It supports multiple platforms including desktop, mobile, and browser environments to enable cloud-edge collaboration and real-time interaction.
- Achieves state-of-the-art results on over 20 GUI benchmarks among open-source models, scoring 56.5 on OSWorld, 71.6 on AndroidWorld, and 48.4 on WebArena for automation tasks.
- Reaches 80.3 on ScreenSpotPro for grounding tasks, 47.6 on OSWorld-MCP and 46.8 on MobileWorld for tool-calling, and 75.5 on GUI-Knowledge Bench for memory and knowledge tasks.
- Incorporates a Hybrid Data Flywheel using simulated and cloud-based sandbox environments to improve data collection efficiency and quality.
- Utilizes a unified thought-synthesis pipeline to enhance reasoning capabilities, focusing on Tool/MCP use, memory, and multi-agent adaptation.
- Proposes the MRPO environment RL algorithm to address multi-platform conflicts and low training efficiency in long-horizon tasks.
The GUI-Owl-1.5 models are open-sourced, with an online cloud-sandbox demo available.