evaluate on Android World

[XiaokunFeng]

Thanks for your great work. May I ask how you conduct evaluations on Androidworld? Do you use only your own prompts, or do you combine them with AndroidWorld’s prompts (including information about UI elements)?

[junyangwang0410]

Hello. Thank you for your attention. The AndroidWorld benchmark code is now open source at [Link]. We've incorporated some context-sensitive prompts from AndroidWorld into Mobile-Agent-v3, but not all. Otherwise, Mobile-Agent-v3 doesn't use any other benchmark-specific prompts and remains consistent with the non-benchmark version.

For UI element detection, we don't use any information, including but not limited to XML or HTML. All UI detection is handled entirely by GUI-Owl.