MobileWeave: Program-First Synthesis of Dependency-Linked Tasks for Training Mobile GUI Agents
Organizations: The University of Hong Kong · LIGHTSPEED · The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology ·
Abstract
Native mobile GUI agents require large-scale training trajectories covering both individual app functions and multi-goal workflows within and across apps, which makes automatic task synthesis a promising way to scale training data. Existing synthesis pipelines, however, produce each task as free-form text, leaving implicit how its subgoals depend on one another. This matters because composition is a skill in its own right: a policy may complete each subtask of a workflow on its own, yet fail to chain these subtasks into the complete workflow. We present MobileWeave, which instead synthesizes every task as an executable program derived from the app implementations of a controllable simulator. MobileWeave abstracts each app into a typed capability graph with semantic types shared across apps, composes atomic programs into single- and cross-app composite programs through explicit value and object dependencies, and verifies each program by native execution, deriving both its instruction and its completion check from the same program. Trained on teacher trajectories collected for these tasks and on recovery supervision at the policy's own stalls, MobileWeave-8B improves its base model by 26.7 points on MobileGym and by 8.6, 21.3, and 17.6 points on AndroidWorld, AndroidLab, and MobileWorld, with comparable gains at 4B. Beyond these gains, we find that composition is learnable from task structure: explicit dependencies between subgoals act as a training signal that teaches a policy to compose skills it already masters. We release data, model, and code at https://maxxnotmarx.github.io/MobileWeave/ .
Figures & tables
| Model | Base Model | Task OOD | Env OOD | |||
|---|---|---|---|---|---|---|
| MobileGym | AndroidWorld | AndroidLab | MobileWorld | |||
| SR | PR | |||||
| Frontier Models | ||||||
| GPT-6.1-Sol [ 13 ] | – | 76.2 | 85.3 | 90.5 | 75.4 | 76.9 |
| DeepSeek-V4-Flash [ 5 ] | – | 47.7 | 59.7 | 57.8 | 50.0 | 58.1 |
| Qwen3.8-Max [ 16 ] | – | 62.9 | 71.7 | 85.3 | 71.0 | 77.8 |
| Training data | #Traj. | Objective | Scope | Overall | |
|---|---|---|---|---|---|
| Single | Hybrid | Cross-app | |||
| (218) | (38) | (93) | |||
| MobileWeave (A + C) | 600 | 38.0 | 38.2 | 12.8 | 38.0 |
| Task source | |||||
| Direct (A + C) | 600 | 30.3 | 22.4 | 0 8.1 | 29.2 |
| Composition | |||||
| SR (%) | Action Loop 3 | |||
|---|---|---|---|---|
| Benchmark | A + C | + Rec. | A + C | + Rec. |
| MobileGym | 40.8 | 44.3 +3.5 | 95.3 | 75.0 21% |
| AndroidWorld | 53.4 | 56.0 +2.6 | 12.0 | 6.7 44% |
| AndroidLab | 59.2 | 64.5 +5.3 | 15.0 | 10.3 31% |
| Training data | Split (42) | Other (211) |
|---|---|---|
| Atomic only | 9.2 | 83.7 |
| + 200 independent | 40.8 | 88.9 |
| + 200 dependency-linked | 60.1 | 92.1 |
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Semantic type | Description | App count | App coverage |
|---|---|---|---|
| Boolean | Truth value of a specific proposition. | 16 | 12306, Alipay, Calendar, Clock, Map, Notes, Redbook, Reddit, Settings, SMS, Spotify, Tencent Meeting, Weather, WeChat, WeChat Reading, X |
| LanguageText | Natural-language text, such as a title or message body. | 15 | 12306, Alipay, Calendar, Clock, eBay, File Manager, Notes, Redbook, Reddit, SMS, Spotify, Tencent Meeting, WeChat, WeChat Reading, X |
| DateTimeText | Displayed date and time text, potentially with incomplete date or time precision. | 9 | 12306, Alipay, Calendar, Clock, Notes, Reddit, SMS, Tencent Meeting, X |
| PersonName | Display name of a person; identical names do not establish identity. | 8 | 12306, Alipay, Redbook, Reddit, Tencent Meeting, WeChat, WeChat Reading, X |
| Count | Nonnegative number of elements in a collection. | 7 | 12306, Alipay, Calendar, Redbook, Reddit, Spotify, Tencent Meeting |
| Region | Region description, potentially at different administrative levels. | 6 | Clock, eBay, Redbook, Weather, WeChat, X |
| Shape | Atomic | Edges | Example | Tasks | Cross-app |
| Q O | 2 | value | Read the school year of a 12306 student discount; set it as the note of an Alipay bill | 342 | 173 |
| Q O | 2 | object | Find the latest event in a date range; delete it | 110 | 0 |
| Q Q | 2 | object | Find a contact; count their chats with unread messages | 20 | 0 |
| O O | 2 | object | Rename a file; move it to a folder | 114 | 0 |
| O Q | 2 | state | Switch the wind-speed unit; read a city’s wind speed | 3 | 0 |
| O, O | 2 | none | Turn off severe-weather alerts; set the pressure unit | 195 | 101 |
| MobileGym test instruction | Synthesized training instruction | Sim. | Difference |
|---|---|---|---|
| Search eBay for electric fans, find the cheapest one, and record its title and price in Notes. | In WeChat, check the region of contact Liu Lang; in eBay, search for electric fans, filter items by the contact’s continent, and get the brand of the item with the lowest total price including shipping; create a note in Notes with this brand as the title and “Tidy up the desk on Saturday morning, take a walk in the afternoon.” as the body. | 0.839 | Different filter, ranking key, and recorded field; an additional WeChat query |
| Change the default advance reminder of Calendar to 5 minutes before. | In Calendar settings, set the snooze time to 5 minutes later. | 0.859 | Different setting |
| Turn off the 07:00 alarm. | Delete the 04:30 alarm labeled “Pick up stuff”. | 0.853 | Different operation and alarm |
| In 12306, add the invoice title Zhang Wei, set it as default, and set the invoice email to [email protected]. | In 12306, set the invoice email to [email protected]. | 0.852 | Different value; no invoice title |
| Set the theme of Map to light. | Set the theme of eBay to light. | 0.810 | Different app |
Mobile GUI Agents
Mobile GUI agents use visual and structured interface information to carry out tasks across smartphone apps.
See all 91 papers