UI-Venus-2 Technical Report
Venus Team, Zhuohan Cai, Haoxing Chen, Jiaxuan Chen, Weizhi Chen, Changlong Gao, Zhangxuan Gu, Yuan Guo and 23 others
Venus Team, Ant Group
Published August 27, 2026 · on the feed September 2, 2026 · ▲ 59 · arXiv 2609.00028
The abstract
Multimodal GUI agents have emerged as a promising paradigm for digital task automation, yet transitioning from benchmark-oriented models to dependable real-world applications remains challenging due to limited environment coverage, brittle task construction, and unreliable reward verification. In this work, we present UI-Venus-2, a general-purpose foundation GUI agent designed to operate across mobile, web, and desktop environments through a unified closed-loop reasoning-action framework. To bridge the gap toward practical deployment, we jointly scale three critical dimensions: (1) Environments, expanding coverage to more than 170 multilingual mobile apps and native desktop operating systems; (2) Tasks, employing a deep-research pipeline for function-grounded instruction generation; and (3) Verification, adopting trace-level and sample-level evaluators with visual keypoints and multi-model voting to ensure reliable RL signals for training. Furthermore, we integrate safety-aware mechanisms to ensure controlled execution of consequential actions. By offering a capable, efficient, and open-source foundation, UI-Venus-2 advances the field toward more generalizable, verifiable, and self-reflective agents for real-world applications.
The Brief
2:13A short podcast summary of the paper. Written and voiced by AI.
Transcript
UI-Venus-2 Technical Report. By Venus Team and thirty others, from the Venus Team at Ant Group. Published on the twenty-seventh of August. Its code repository is worked out: about fifty-five source files, with a readme.
Multimodal GUI agents, models that carry out tasks by operating a graphical interface, are a promising paradigm for automating digital work. Moving them from benchmark-oriented models to dependable real-world applications is still hard, and the authors name three reasons. Environment coverage is limited. Task construction is brittle. And reward verification is unreliable, which matters because those rewards are the signal that reinforcement learning trains on.
UI-Venus-2 is a general-purpose foundation GUI agent built to operate across mobile, web and desktop environments, through a unified closed-loop reasoning-action framework. One agent, then, rather than a separate model for each kind of device. Its response to the three problems is to scale all three critical dimensions jointly, environments, tasks and verification, with practical deployment as the target. Each of the three shortcomings gets a matching line of work.
For environments, coverage expands to more than one hundred and seventy multilingual mobile apps, along with native desktop operating systems. For tasks, a deep-research pipeline handles function-grounded instruction generation. For verification, evaluators work at both the trace level and the sample level, using visual keypoints and voting across multiple models, so that the RL signals used in training are reliable. Several judges looking at both whole traces and individual samples is the answer to the third complaint, rewards that cannot be trusted. On top of all this, safety-aware mechanisms are integrated to keep the execution of consequential actions under control.
The abstract reports no benchmark numbers. What it claims is the offering itself: a foundation that is capable, efficient and open-source.
The direction set here is toward agents that are more generalizable, more verifiable and more self-reflective in real-world use. The bet behind UI-Venus-2 is that the distance from benchmark to deployment closes by scaling the conditions an agent learns under: wider environments, sturdier tasks, and rewards that can be trusted.
Paper Quest · 7 chapters · about 33 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Benchmark scores look good. Real phones and desktops disagree. Start there.
You are building an agent that operates phones, websites and desktops by looking at the screen and clicking, typing and scrolling. It does well on benchmarks, then stumbles in real use.
The road ahead
- 1 · The problem ~7 min A screen-operating agent that scores well on benchmarks still stumbles in real use, because its apps, tasks and reward checker are all too weak.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Make the agent dependable by scaling environments, tasks and reward verification together, since a weakness in any one spoils the other two.
- 4 · Choose your own research ~3 min
- 5 · How it works ~3 min
- 6 · Field test ~3 min
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.