Reproduce “UI-Venus-2 Technical Report” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Venus Team, Zhuohan Cai, Haoxing Chen, Jiaxuan Chen, Weizhi Chen, Changlong Gao, Zhangxuan Gu, Yuan Guo and 23 more
UI-Venus-2 expands a closed-loop GUI agent across mobile, web, and desktop. It builds executable tasks from capability catalogs and verifies progress at trajectory and action levels. On VenusBench-Mobile, its 27B model completes 48.7% of 149 tasks, versus 36.5% for Claude Opus 4.6; this is benchmark evidence, not a measure of deployment reliability.
Key insights
- Jointly scaling environment diversity, task distribution complexity, and data verification reliability is crucial for developing general-purpose GUI agents.
- Structured action-aware distillation and teacher-side action-type conditioning in MOPD effectively integrate specialized skills and broad multimodal reasoning by focusing distillation signals on executable behaviors.
- A closed-loop data generation framework, iteratively refining capability catalogs, synthesizing tasks, and collecting trajectories with domain-specific processes, ensures high-quality and reliable training data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.