UI-Venus-2 Technical Report: An Open-Source Multimodal GUI Agent

Summary

UI-Venus-2 is an open-source multimodal GUI agent designed for task automation across mobile, web, and desktop environments. Its training approach expands environment coverage, grounds task instructions in functions, and uses visual keypoints, trace-level evaluation, sample-level evaluation, and multi-model voting to improve reinforcement-learning signals. Safety-aware mechanisms help control consequential actions.