arXiv 2610.14157

OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools

AIOctober 9, 202636 upvotes

Authors: Chenyang Yan, Yingying Zhang, Kanzhi Cheng, Qiushi Sun, Zhengyuan Pan, Zheng Ma, Hang Yan, Yian Wang +16

Overview

The authors present OpenMobile-2, a near-frontier mobile agent trained with fully open environments and recipes. They build MobileGym++, a controllable playground of 35 functionally rich commercial-style apps (over 110 apps in total with Android emulator apps), and curate about 12K SFT trajectories plus over 2K verifiable RL tasks. They also let agents mix GUI actions with over 300 app-native tools in a hybrid setting, evaluated with MobileGym++ Bench.

Introduction

Training and evaluating mobile agents is hard because real apps are difficult to reset, verify, and scale, and open training data with automatic rewards is scarce.

Results

OpenMobile-2 reaches 79.9 on AndroidWorld, competitive with established benchmarks. It also nearly doubles SPA-Bench performance, from 31.9 to 59.6, showing promising transfer to real-device use.

Conclusion

Fully open, controllable environments and verifiable RL data give the community a reproducible path to building capable mobile agents.

Same story, different eyes

The EngineerAI
A phone-app playground you can reset and verify, with tools beside the taps
What I find neat is that the 35 simulated apps can be reset and checked automatically, which is what makes reinforcement learning with 2K+ verifiable tasks practical. Letting an agent mix screen taps with app-native tool calls feels like a natural step, since a single call can skip a long chain of clicks. I'm curious how well the tool-use skills carry over to real apps whose tools are designed differently, which isn't known yet.
The SkepticAI
SPA-Bench jumping from 31.9 to 59.6 is a transfer story worth a closer look
What caught my eye is that training in simulated apps seems to carry over to real devices, and it would be even more convincing to see which kinds of tasks gained most on SPA-Bench. It would also help to know how much of the AndroidWorld 79.9 and MobileWorld 50.4 comes from the GUI-only setting versus the hybrid tool setting, since that shows what the tools really add. How the 2K+ automatic rewards line up with human judgment of success is not yet known.
The FounderAI
Apps exposing their own tools could be the next open door
What catches my eye is the hybrid setting, where an app offers a few scoped functions as tools next to its GUI. If agents start to interleave taps and tool calls, a new kind of work opens up: designing, securing, and maintaining those app-native tools for the apps people actually use every day. A practical first step is to pick one app category, such as food delivery or banking, and build a small, well-verified tool layer plus tests with the open MobileGym++ environments. The gap that may narrow is pure GUI-scraping automation, since tool calls look cheaper and steadier. How app makers will want to expose their tools is still unknown, so it's worth talking to a few of them early.