Summary
UI-Venus-2 (arXiv 2609.00028, Wed 2 Sep digest, 31 authors under the Venus Team banner) is a general-purpose, open-source foundation GUI agent spanning mobile, web, and desktop through a unified closed-loop reasoning-action framework. It scales three axes jointly: environments (170+ multilingual mobile apps plus native desktop operating systems), tasks (a deep-research pipeline generating function-grounded instructions), and verification (trace-level and sample-level evaluators combining visual keypoints and multi-model voting, to make RL training signals reliable). It adds safety-aware mechanisms around consequential actions and is positioned as an open foundation for more generalizable, verifiable, self-reflective agents.
Why it matters
Computer-use agents live or die on the reliability of their RL signal, and the hardest part is verifiable reward at trajectory level. This release bundles environment, task generation, and a two-level verifier design in one open package — a reference point for anyone building self-hosted GUI agents, and the most complete open recipe in the computer-use direction since the category became a focus area.
Technical details
| Arxiv | 2609.00028, announced in the Wed 2 Sep 2026 digest (submitted 8/27) |
|---|---|
| Scope | general-purpose foundation GUI agent: mobile + web + desktop via a unified closed-loop reasoning-action framework |
| Environments | 170+ multilingual mobile apps plus native desktop operating systems |
| Tasks | deep-research pipeline generating function-grounded instructions |
| Verification | trace-level and sample-level evaluators: visual keypoints + multi-model voting, aimed at reliable RL signals |
| Safety | safety-aware mechanisms around consequential actions |
| Openness | described as open-source in the abstract; no explicit repo URL on the arXiv abstract page at capture time |
Tags
computer-usegui-agentopen-sourcerlverificationagent