Most GUI agent research focuses on benchmarks; the last step toward real-world autonomy is robust coverage, executable tasks, and reliable verification. UI-Venus-2 tackles these three failure modes jointly by enlarging environment pools, grounding generated instructions in actual app functionality, and replacing coarse completion checks with keypoint-driven, multi-model voting verification to yield dependable RL signals.
Key Findings
- Broad environment scaling: executable pools include 170+ multilingual mobile apps (Chinese + English), a curated web pool of 4,000+ domains across many categories, and native desktop OS support — this increases encounter diversity and reduces brittle task assumptions.
- Task grounding and pipeline: function-grounded query generation produces executable, internally consistent task instructions, lowering invalid/ambiguous trajectories in training data.
- Trace- and sample-level verification: completion judged against task-relevant visual keypoints and aggregated across heterogeneous models (voting), improving robustness to superficial progress and reward-hacking.
- Three-stage training recipe: large-scale multimodal mid-training (trajectory injection), step-level offline RL per domain (Grounding, CAPTCHA, Mobile, Web, Computer), and multi-teacher on-policy distillation to consolidate domain-specialized policies into a unified model.
- Safety-aware controls: targeted CAPTCHA handling and consequential-action safety evaluation; reported safety metrics for a 9B checkpoint include OSHarm ASR 11.3% and OSBlind ASR 48.8%, improving over the Qwen3.5-9B baseline.
Who it's for & tradeoffs
Great fit if you need a foundation GUI policy for autonomous interface tasks that must generalize across mobile, web, and desktop, and you value verified RL signals and reproducible weights/benchmarks. It suits research groups building end-to-end GUI agents, teams collecting verified GUI interaction data, or projects needing an off-the-shelf multimodal agent for automation.
Look elsewhere if you require lightweight on-device inference (the released family includes large checkpoints and assumes server-grade compute), strictly private closed-source deployment without auditing (weights and infra are publicly released under a repository workflow), or narrower single-domain solutions where a specialized, smaller model may be cheaper to run.