OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools
Chenyang Yan, Yingying Zhang, Kanzhi Cheng, Qiushi Sun, Zhengyuan Pan, Zheng Ma, Hang Yan, Yian Wang, Nuo Chen, Jialin Cao, Xingdong Gong, Wenpo Song, Han Chen, Zichen Ding, Fangzhi Xu, Shujian Huang, Xinyu Dai, Yichen Liu, Tiankuo Yao, Bo Wang, Ben Kao, Jianbing Zhang, Lewei Lu, Dahua Lin
cs.SE, cs.CL, cs.CV
2026-10-10
OpenMobile-2 releases a 110-app playground, 11.6K traces, and 2,475 verifiable RL tasks. The 27B model hits 79.9 on AndroidWorld and real-device SPA-Bench of 59.6, up from 31.9.
Open mobile agents are short of a place to practice. Taps and typing are learned from cross-app trajectories, and reinforcement learning needs those runs to reset and to mark success. Real apps in an Android emulator look like everyday phone use, but logins and irreversible side effects make large parallel training expensive. Simulated clients reset cleanly, yet the apps are few and the flows are shallow. The previous MobileGym included 12 commercial-style apps.
OpenMobile-2, from Nanjing University, SenseTime, and collaborators, scales that playground. MobileGym++ adds 35 commercial-style apps. Together with newly configured emulator apps, the pool exceeds 110. The release has 11,643 demonstration trajectories and 2,475 tasks with automatic rewards. Apps can also expose their own functions, which the agent interleaves with taps.
MobileGym++ runs in the browser. The agent sees pixels only: no accessibility tree, and no direct read of orders or itineraries. Device state can be saved and restored, so a reset is deterministic. The 35 apps were specified from real products' data schemas and interaction graphs before the UI was drawn, then edited by people against the live apps. Every declared interaction had an implementation, and core flows were replayed against the promised state change. The median app has 166 interactive actions, and median depth from the home screen is 3. Against the 12 older commercial apps kept in the same environment, the new apps have 1.48 times the actions and 1.91 times the page transitions.
Cross-app work uses the clipboard, sharing, payment handoff, and media picking on one resettable device. The simulator side has 63 apps, 35 of them new; the emulator side has 62.
App-native tools write the same business objects as taps and face the same success check, so the two channels can be compared on one task list. A call such as adding a product to the cart by id is one structured read or write. More than 300 tools sit on 50-plus apps, and a catalog is visible only while that app is in the foreground. All 215 benchmark tasks remain solvable with taps alone. On the Taobao replica, prepare-checkout only opens a confirmation page and does not create an order.
The hybrid training set has 11,643 trajectories: 7,725 from the emulator and 3,918 from simulated apps, including 1,774 that mix taps and tool calls. Each of the 2,475 RL tasks stores a goal, an initial state, and a verifier.
Tasks come from a state-action graph. Emulator graphs are built by exploration and by merging similar screens; simulated apps read their backend spec. Gemini-3.1-Pro-Preview writes the user goals and then executes them. Goals do not name widgets. Emulator traces are filtered by a vision-language judge, simulated traces by a program, and steps that leave state unchanged are dropped. 30% of trajectories cross apps.
The bases are Qwen3.5-9B and Qwen3.6-27B. Each gets two supervised runs of 5 epochs, one on GUI traces and one on the mixture, keeping the last 3 screenshots. RL continues from the hybrid checkpoint with GRPO-style group comparisons and DAPO's asymmetric clipping, on 512 randomly chosen tasks with 8 samples each. The vision tower stays frozen and only a rank-64 LoRA is trained. The training reward uses a relaxed check, meaning the goal is met and extra changes are benign, then multiplies by 2 if the trajectory called an app-native tool at least once. Reported accuracy uses a strict check that demands a clean state.
MobileGym++ Bench has 215 tasks: 109 on one app, 74 on two, 27 on three, and 5 on four. The checker reads the business-state diff and ignores screenshots.
The 27B line starts from Qwen3.6-27B. Bench means MobileGym++ Bench.
| Stage | AndroidWorld | MobileWorld GUI | MobileWorld MCP | MobileGym | Bench GUI | Bench hybrid | SPA-Bench |
| Base | 67.1 | 35.9 | 32.5 | 37.1 | 46.5 | 50.7 | 31.9 |
| GUI SFT | 80.8 | 50.4 | 35.0 | 46.9 | 54.9 | 57.7 | not reported |
| Hybrid SFT | 80.8 | 48.7 | 45.0 | 48.4 | 55.3 | 65.6 | 57.4 |
| Plus RL | 79.9 | 50.4 | 42.5 | 51.2 | 57.2 | 67.0 | 59.6 |
On AndroidWorld the gain is from supervised training: 67.1 to 80.8, then 79.9 after RL. Pass@3 moves from 81.9 to 90.5. The 9B model moves from 54.3 to 74.0. Reported AndroidWorld scores include 76.7 for MAI-UI-235B, 70.7 for the Gemini-3.1-Pro preview, and 73.7 for UI-Venus-1.5-8B.
On MobileWorld GUI, 50.4 is above MAI-UI-235B at 39.7 and below GPT-5.6-Sol at 70.1 and Gemini at 58.1.
On real-device SPA-Bench L3, the 27B model goes from 31.9 to 57.4 after hybrid SFT and to 59.6 after RL. The 9B model goes from 25.5 to 48.9. The reproduced GPT-5.6-Sol score in that column is 29.8.
After hybrid SFT, turning tools on adds 5.6 points for the 9B model (39.1 to 44.7) and 10.3 for the 27B model (55.3 to 65.6). Without that training, Qwen gains only 2.8 to 4.2. After training, the two models call tools on 113 and 128 of 215 tasks and finish 4.8 and 5.8 steps sooner. On MobileWorld MCP tools that never appeared in training, hybrid SFT scores 27.5 and 45.0, against 12.5 and 35.0 for GUI-only SFT. RL then drops the 27B MCP score from 45.0 to 42.5.
MobileGym success moves from 35.9 to 38.3 at 9B and from 48.4 to 51.2 at 27B. AndroidWorld does not rise further.
With the budget fixed at 2,000 trajectories, widening coverage from 15 apps to 53 raises held-out AndroidWorld Pass@1 from 60.7 to 67.5. In-domain Pass@1 goes from 73.7 to 71.4 while the in-domain share falls from 100% to 39.1%. At most 2.25% of SFT goals exceed cosine similarity 0.7 with a test set, and none exceeds 0.9. Of the 2,475 RL goals, 6.1% exceed 0.7 against MobileGym++ Bench.
The usable release is the whole stack: a resettable environment, 11,643 trajectories, 2,475 auto-graded tasks, a 215-task bench, and the 9B and 27B models. The 35 replicas have 1.48 times the interactive actions of the previous simulated commercial apps, with reset and programmatic checks still in place.
At a fixed 2,000 trajectories, growing the app set from 15 to 53 adds 6.8 points on unseen apps and barely moves the familiar ones. When the data budget is tight, extra apps beat extra repeats of the same apps.
Tools start to matter once they share state and the success check with taps, and once mixed traces are in the training set. Success goes up and episodes get shorter. RL adds 2.4 to 2.8 points on MobileGym and 2.2 on real devices. Other groups can keep training on the 2,475 tasks that already have verifiers.
The commercial apps are replicas, not official clients. The paper reports no measured visual fidelity or flow coverage, and median depth from the home screen is 3.
Goals and demonstrations were produced by Gemini-3.1-Pro-Preview. The data and recipe are open; synthesis still depends on a closed model. Emulator traces were filtered by a vision-language judge, and that label noise is not measured.
One app-native tool call multiplies the training reward by 2, which pushes the policy toward calling tools. Reported scores use a strict checker, a different yardstick from that reward. External comparisons also keep each model's own prompt, so the AndroidWorld ranking against the 235B model is not a same-protocol result.
The real-device move from 31.9 to 59.6 is large, and the reproduced GPT-5.6-Sol number is 29.8, but most strong open models have no SPA-Bench entry in this table. On MobileWorld GUI, 50.4 still sits below GPT-5.6-Sol at 70.1 and Gemini at 58.1. RL barely moves scores outside the MobileGym family, and AndroidWorld for the 27B model slips from 80.8 to 79.9. On the in-house bench, 6.1% of RL goals exceed similarity 0.7 with the test goals.