A Benchmark for Mobile GUI Agent Evaluation · Statistics & Visualization
MobileGUI-VBench is a benchmark suite for evaluating mobile GUI Agent capabilities. It covers a wide range of real-world scenarios including Social & Communication, Music & Radio, Video Streaming, Shopping & Deals, Travel & Transit, Navigation, Lifestyle Services, and News & Reading. The benchmark encompasses single-app and cross-app tasks, single-intent and multi-intent tasks, explicit instructions and implicit-intent instructions, and is designed to evaluate an Agent's instruction comprehension and execution capability in real mobile app environments.
Data files: MobileGUI-VBench.xlsx (full annotations), MobileGUI-VBench.jsonl (reduced field set)
| Column | Description |
|---|---|
| TASK_ID | Unique task identifier |
| Instruction | Raw user query / instruction text |
| Scenario Category | Task scenario category (e.g., Social & Communication, Music & Radio, Video Streaming) |
| APP | App(s) involved in the task, multiple apps separated by 「、」 |
| Instruction Type | Agent-Execution / Agent-User Interaction |
| Chain Complexity | Simple / Medium / Complex |
| Operation Type | Basic Operation / Advanced Operation |
| Instruction Clarity | Explicit Instruction / Implicit Intent Instruction / Ambiguous Instruction |
| Intent Composition | Single Intent / Dependent Multi-Intent / Independent Multi-Intent |
| Function Point | Specific functionality being tested (e.g., Search, Product Lookup, Watch Video) |
| Precondition | Required state or environment setup for task completion |
| Environment Preset | Specific environment setup steps (provided for some tasks) |
| Ideal Clarification | Expected clarification behavior from the Agent (mostly for Agent-User Interaction tasks) |
This section defines the multi-dimensional tags used in MobileGUI-VBench, serving as the criteria for task annotation and result analysis. The tag system consists of 6 classification dimensions (Instruction Type / Chain Complexity / App Scope / Operation Type / Instruction Clarity / Intent Composition) and 2 auxiliary fields (Precondition / Ideal Clarification).
| Tag | Definition | Example |
|---|---|---|
| Agent-Execution | The instruction is complete and unambiguous; the Agent can execute it directly without asking the user | "Set my QQ status to Busy" |
| Agent-User Interaction | The instruction lacks sufficient information or has multiple valid execution paths; the Agent is expected to proactively clarify with the user before executing | "Change my QQ status to something else" (needs clarification: to which status?) |
Categorized by the estimated number of steps required to complete the task (each tap / input / swipe counts as one step):
| Tag | Definition | Example |
|---|---|---|
| Simple | Completable within ≤ 5 steps | "Open NetEase Cloud Music" |
| Medium | Requires 6 ~ 10 steps | "Search and play a tech video on Bilibili" |
| Complex | Requires > 10 steps, or involves deep multi-page navigation | "Find the top-rated coffee shop on RED and save the post" |
| Tag | Definition | Example |
|---|---|---|
| Single-App | The task is completed entirely within one app | "Play my favorite songs in QQ Music" |
| Cross-App | The task requires switching between ≥ 2 apps or transferring data across apps | "Open in Amap the location my friend sent me in WeChat" |
| Tag | Definition | UI Actions Involved |
|---|---|---|
| Basic Operation | Uses only the most common GUI interactions | Tap, text input |
| Advanced Operation | Involves more complex gestures or control manipulation | Swipe, drag, long-press, double-tap, pinch-zoom, scroll-targeting |
| Tag | Definition | Example |
|---|---|---|
| Explicit Instruction | The instruction clearly specifies the target, action, and parameters; no inference required | "Pin my chat with Zhang San in WeChat" |
| Implicit Intent Instruction | The instruction does not explicitly state the real intent; the Agent must infer it from common sense or context | "Haven't been on QQ for a while, see what my friends have been posting" (implicit intent: view friend feeds) |
| Ambiguous Instruction | The instruction lacks key parameters or has multiple valid interpretations; the Agent must clarify | "Help me search for that movie" (which movie? on which platform?) |
| Tag | Definition | Example |
|---|---|---|
| Single Intent | The instruction contains only one independent task | "Open NetEase Cloud Music and play my favorite playlist" |
| Dependent Multi-Intent | The instruction contains multiple tasks with conditional dependency or data passing between them (B requires the result of A) | "Search for hotpot restaurants near me on Amap, then send the nearest one to Zhang San on WeChat" (B depends on A's output) |
| Independent Multi-Intent | The instruction contains multiple tasks that are independent and can be executed in parallel | "Set my QQ status to Busy, and also open NetEase Cloud Music" |
The environment state or account state required before executing the task. An empty field means no special precondition is needed.
Examples:
For Agent-User Interaction tasks, this field defines the clarification or confirmation behavior the Agent is expected to proactively initiate. It is the core basis for evaluating the Agent's proactive interaction capability.
Examples:
| Scenario Category | Count | Percentage |
|---|---|---|
| Video Streaming | 24 | 16.0% |
| Shopping & Deals | 24 | 16.0% |
| Lifestyle Services | 23 | 15.3% |
| Social & Communication | 21 | 14.0% |
| Travel & Transit | 18 | 12.0% |
| Music & Radio | 17 | 11.3% |
| Navigation | 17 | 11.3% |
| News & Reading | 6 | 4.0% |
The dataset involves 40 unique apps. A single task may involve multiple apps, so the total count of app occurrences exceeds the number of tasks (150).
All 40 apps shown, sorted by occurrence frequency (descending).
| Instruction Type | Count | Percentage |
|---|---|---|
| Agent-Execution | 91 | 60.7% |
| Agent-User Interaction | 59 | 39.3% |
| Chain Complexity | Count | Percentage |
|---|---|---|
| Medium | 90 | 60.0% |
| Simple | 36 | 24.0% |
| Complex | 24 | 16.0% |
| Operation Type | Count | Percentage |
|---|---|---|
| Basic Operation | 123 | 82.0% |
| Advanced Operation | 27 | 18.0% |
| Instruction Clarity | Count | Percentage |
|---|---|---|
| Explicit Instruction | 76 | 50.7% |
| Implicit Intent Instruction | 67 | 44.7% |
| Ambiguous Instruction | 7 | 4.7% |
| Intent Composition | Count | Percentage |
|---|---|---|
| Single Intent | 93 | 62.0% |
| Dependent Multi-Intent | 49 | 32.7% |
| Independent Multi-Intent | 8 | 5.3% |
The function points are diverse (77 types in total). Only the top 15 by frequency are shown here. The remaining function points mostly appear 1–2 times, covering more granular long-tail operation scenarios.
The MobileGUI-VBench dataset contains 150 tasks across 40 mainstream mobile apps and 8 real-world scenario categories. The task design balances single-app and cross-app collaboration, single-intent and multi-intent workflows, and explicit, implicit, and ambiguous instruction comprehension, making it a comprehensive benchmark for evaluating the scenario generalization and complex instruction execution capabilities of mobile GUI Agents.