OpenMobile-2 Advances Mobile Agents with Open Environments and Native Tools

Introduction

A mobile agent that can complete tasks such as coordinating information across apps or following a long workflow needs more than a capable language model. Researchers also need diverse applications, repeatable initial states, reliable task verification, and training data that can support both supervised learning and reinforcement learning. OpenMobile-2 focuses on assembling these pieces into an open research stack.

Key contributions

• A scalable environment for mobile interaction. The project introduces MobileGym++, a collection of 35 realistic, functionally rich commercial-style apps. The apps support cross-application workflows while preserving control over resets and outcome verification. Combined with configured apps in Android emulators, the playground spans more than 110 apps.
• Training data with executable rewards. The authors curate nearly 12,000 mobile interaction trajectories for supervised fine-tuning. They also release more than 2,000 executable tasks for verifiable reinforcement learning, with automatic rewards that can check whether a task has been completed. This reduces dependence on manual inspection during large-scale training.
• Hybrid GUI and native-tool use. Instead of forcing an agent to rely exclusively on taps, swipes, and text entry, selected app functions can be exposed as scoped native tools. An agent may alternate between GUI actions and tool calls within one task. MobileGym++ implements more than 300 such tools across over 50 apps.
• A shared long-horizon benchmark. MobileGym++ Bench targets complex commercial-style workflows and supports both GUI-only and hybrid GUI-tool evaluation on the same task suite, making the two interaction styles easier to compare.

Results and implications

OpenMobile-2 reports scores of 79.9 on AndroidWorld and 50.4 on MobileWorld. On SPA-Bench real-device mobile tasks, its score rises from 31.9 to 59.6, nearly doubling the reported performance. These results suggest that better environments, verifiable data, and flexible action interfaces can be as important as model architecture when improving mobile task completion.

The broader contribution is an attempt to open the full experimental loop: environments, data, training recipes, benchmarks, and models are presented as connected resources. This can lower the barrier for researchers who want to reproduce results or create new workflows. The native-tool direction is particularly notable because it avoids a false choice between general visual interaction and efficient API-style execution. Agents can use the GUI where visual context matters and invoke a tool where a narrowly defined function is more reliable.

There are still important boundaries. Simulated commercial-style apps do not fully reproduce the variability of real applications, and hybrid tools introduce questions about permissions, observability, and fair comparison with GUI-only agents. Long-horizon tasks can also accumulate small errors across multiple apps. Real-device transfer results are encouraging, but broader testing will be needed to establish how robust the approach is outside the controlled environment.

Source: Hugging Face Daily Papers