> For AI agents: the complete documentation index is available at /zh/llms.txt, the full documentation bundle is available at /zh/llms-full.txt.
# Midscene AndroidWorld Benchmark 测试报告
本文是 Midscene 针对 AndroidWorld benchmark 的测试报告。本次测试中，Midscene 取得了 **Pass\@1 93.10%**、**Pass\@2 95.69%**、**Pass\@3 97.41%** 的结果。
:::info 关于 AndroidWorld
[AndroidWorld](https://github.com/google-research/android_world) 是 Google Research 提供的 Android Agent benchmark。它运行在真实 Android 模拟器上，用 20 个真实 Android 应用中的 116 个程序化任务来评估 Agent，并由 benchmark 负责初始化任务和校验结果。
:::
## 运行配置
| 字段 | 值 |
| --------------------------------- | ------------------------------------------------------------- |
| Model Name | `Gemini-3.5-Flash` |
| Midscene version | `1.9.5` |
| DeepThink | on |
| `MIDSCENE_REPLANNING_CYCLE_LIMIT` | 120 |
| AndroidWorld 工程 | 为了降低 benchmark 过程中的偶发波动，我们对 AndroidWorld 工程做了若干稳定性修复。具体例子见下方。 |
| 验收规则 | 少量 AndroidWorld validator 按任务意图做了校准，涉及的 case 见下方列表。 |
## 稳定性提升
以下改动不改变任务意图，主要用于降低浏览器渲染、accessibility tree 刷新、任务 setup 时序等带来的偶发失败。
| 修改内容 | 涉及的 case |
| -------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| BrowserDraw 绘制后强制刷新 canvas 像素，并加粗、圆角化画笔，让目标颜色更稳定地落到最终 canvas 像素里。 | `BrowserDraw` |
| Browser task 完成后，读取 accessibility tree 中的 `Success!` 文本时增加重试，避免 stale accessibility tree 误判。 | `BrowserMaze`
`BrowserMultiply`
`BrowserDraw` |
| SMS 任务开始前会先准备好需要的短信和联系人；现在会确认这些短信能在收件箱查到、联系人能在通讯录查到，再开始让 agent 执行。 | `SimpleSmsReplyMostRecent`
`SimpleSmsSendReceivedAddress` |
| Expense 类任务启动 Pro Expense 后等待数据库表创建完成，再由 validator 写入测试数据，避免初始数据库未初始化导致失败。 | `ExpenseAddMultiple`
`ExpenseAddMultipleFromGallery`
`ExpenseAddMultipleFromMarkor`
`ExpenseAddSingle`
`ExpenseDeleteDuplicates`
`ExpenseDeleteDuplicates2`
`ExpenseDeleteMultiple`
`ExpenseDeleteMultiple2`
`ExpenseDeleteSingle` |
| OsmAnd 任务提前把离线地图文件放入应用数据目录，并等待 OsmAnd 解压内置底图，确保地图数据 ready 后再执行。 | `OsmAndFavorite`
`OsmAndMarker`
`OsmAndTrack` |
## 验证条件调整
本次 benchmark 使用的 AndroidWorld `main` 分支中，以下验证条件做过调整：
| 修改内容 | 涉及的 case |
| ---------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Calendar 的 "after start time" 验收目标改为边界点之后 1 分钟的事件，避免事件正好落在边界点时，不同模型对 `after` 是否包含边界点的理解差异影响判定。 | `SimpleCalendarFirstEventAfterStartTime` |
| Expense 从 Markor 导入时，接受 Markor note 中额外带上的 `Reimbursable.` 后缀；对比 note 文本时会忽略这个后缀及末尾句号差异。 | `ExpenseAddMultipleFromMarkor` |
| Markor 合并 notes 时，允许单换行或空行分隔，并兼容 Markor 自动补上的 `.md` 扩展名。 | `MarkorMergeNotes` |
| Recipe 的数量、时长和 ingredient count 允许省略单位，但错误数量或错误单位仍会失败。 | `RecipeAddSingleRecipe`
`RecipeAddMultipleRecipes`
`RecipeAddMultipleRecipesFromMarkor`
`RecipeAddMultipleRecipesFromMarkor2`
`RecipeAddMultipleRecipesFromImage`
`NotesRecipeIngredientCount` |
| 最小亮度按 Android 实际 setting 值 `0` 验证，而不是旧的 `1`。 | `SystemBrightnessMin`
`SystemBrightnessMinVerify` |
## 报告文件
详细的报告如下，供大家参考。
Round 1（115 份报告 · 108 PASS · 7 FAIL）
| # | 任务 | 状态 | 报告 |
| --- | ------------------------------------------------- | -------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 1 | AudioRecorderRecordAudio | PASS | 报告 |
| 2 | AudioRecorderRecordAudioWithFileName | PASS | 报告 |
| 3 | BrowserDraw | PASS | 报告 |
| 4 | BrowserMaze | PASS | 报告 |
| 5 | BrowserMultiply | PASS | 报告 |
| 6 | CameraTakePhoto | PASS | 报告 |
| 7 | CameraTakeVideo | PASS | 报告 |
| 8 | ClockStopWatchPausedVerify | PASS | 报告 |
| 9 | ClockStopWatchRunning | PASS | 报告 |
| 10 | ClockTimerEntry | PASS | 报告 |
| 11 | ContactsAddContact | PASS | 报告 |
| 12 | ContactsNewContactDraft | PASS | 报告 |
| 13 | ExpenseAddMultiple | PASS | 报告 |
| 14 | ExpenseAddMultipleFromGallery | PASS | 报告 |
| 15 | ExpenseAddMultipleFromMarkor | FAIL | 报告 |
| 16 | ExpenseAddSingle | PASS | 报告 |
| 17 | ExpenseDeleteDuplicates | PASS | 报告 |
| 18 | ExpenseDeleteDuplicates2 | PASS | 报告 |
| 19 | ExpenseDeleteMultiple | PASS | 报告 |
| 20 | ExpenseDeleteMultiple2 | PASS | 报告 |
| 21 | ExpenseDeleteSingle | PASS | 报告 |
| 22 | FilesDeleteFile | PASS | 报告 |
| 23 | FilesMoveFile | PASS | 报告 |
| 24 | MarkorAddNoteHeader | PASS | 报告 |
| 25 | MarkorChangeNoteContent | PASS | 报告 |
| 26 | MarkorCreateFolder | PASS | 报告 |
| 27 | MarkorCreateNote | PASS | 报告 |
| 28 | MarkorCreateNoteAndSms | PASS | 报告 |
| 29 | MarkorCreateNoteFromClipboard | PASS | 报告 |
| 30 | MarkorDeleteAllNotes | PASS | 报告 |
| 31 | MarkorDeleteNewestNote | PASS | 报告 |
| 32 | MarkorDeleteNote | PASS | 报告 |
| 33 | MarkorEditNote | PASS | 报告 |
| 34 | MarkorMergeNotes | PASS | 报告 |
| 35 | MarkorMoveNote | PASS | 报告 |
| 36 | MarkorTranscribeReceipt | PASS | 报告 |
| 37 | MarkorTranscribeVideo | FAIL | 报告 |
| 38 | OpenAppTaskEval | PASS | 报告 |
| 39 | OsmAndFavorite | PASS | 报告 |
| 40 | OsmAndMarker | FAIL | 报告 |
| 42 | RecipeAddMultipleRecipes | PASS | 报告 |
| 43 | RecipeAddMultipleRecipesFromImage | FAIL | 报告 |
| 44 | RecipeAddMultipleRecipesFromMarkor | PASS | 报告 |
| 45 | RecipeAddMultipleRecipesFromMarkor2 | PASS | 报告 |
| 46 | RecipeAddSingleRecipe | PASS | 报告 |
| 47 | RecipeDeleteDuplicateRecipes | PASS | 报告 |
| 48 | RecipeDeleteDuplicateRecipes2 | FAIL | 报告 |
| 49 | RecipeDeleteDuplicateRecipes3 | FAIL | 报告 |
| 50 | RecipeDeleteMultipleRecipes | PASS | 报告 |
| 51 | RecipeDeleteMultipleRecipesWithConstraint | PASS | 报告 |
| 52 | RecipeDeleteMultipleRecipesWithNoise | PASS | 报告 |
| 53 | RecipeDeleteSingleRecipe | PASS | 报告 |
| 54 | RecipeDeleteSingleWithRecipeWithNoise | PASS | 报告 |
| 55 | RetroCreatePlaylist | PASS | 报告 |
| 56 | RetroPlayingQueue | PASS | 报告 |
| 57 | RetroPlaylistDuration | PASS | 报告 |
| 58 | RetroSavePlaylist | PASS | 报告 |
| 59 | SaveCopyOfReceiptTaskEval | PASS | 报告 |
| 60 | SimpleCalendarAddOneEvent | PASS | 报告 |
| 61 | SimpleCalendarAddOneEventInTwoWeeks | PASS | 报告 |
| 62 | SimpleCalendarAddOneEventRelativeDay | PASS | 报告 |
| 63 | SimpleCalendarAddOneEventTomorrow | PASS | 报告 |
| 64 | SimpleCalendarAddRepeatingEvent | PASS | 报告 |
| 65 | SimpleCalendarDeleteEvents | PASS | 报告 |
| 66 | SimpleCalendarDeleteEventsOnRelativeDay | PASS | 报告 |
| 67 | SimpleCalendarDeleteOneEvent | PASS | 报告 |
| 68 | SimpleDrawProCreateDrawing | PASS | 报告 |
| 69 | SimpleSmsReply | PASS | 报告 |
| 70 | SimpleSmsReplyMostRecent | PASS | 报告 |
| 71 | SimpleSmsResend | PASS | 报告 |
| 72 | SimpleSmsSend | PASS | 报告 |
| 73 | SimpleSmsSendClipboardContent | PASS | 报告 |
| 74 | SimpleSmsSendReceivedAddress | PASS | 报告 |
| 75 | SystemBluetoothTurnOff | PASS | 报告 |
| 76 | SystemBluetoothTurnOffVerify | PASS | 报告 |
| 77 | SystemBluetoothTurnOn | PASS | 报告 |
| 78 | SystemBluetoothTurnOnVerify | PASS | 报告 |
| 79 | SystemBrightnessMax | PASS | 报告 |
| 80 | SystemBrightnessMaxVerify | PASS | 报告 |
| 81 | SystemBrightnessMin | PASS | 报告 |
| 82 | SystemBrightnessMinVerify | PASS | 报告 |
| 83 | SystemCopyToClipboard | FAIL | 报告 |
| 84 | SystemWifiTurnOff | PASS | 报告 |
| 85 | SystemWifiTurnOffVerify | PASS | 报告 |
| 86 | SystemWifiTurnOn | PASS | 报告 |
| 87 | SystemWifiTurnOnVerify | PASS | 报告 |
| 88 | TurnOffWifiAndTurnOnBluetooth | PASS | 报告 |
| 89 | TurnOnWifiAndOpenApp | PASS | 报告 |
| 90 | VlcCreatePlaylist | PASS | 报告 |
| 91 | VlcCreateTwoPlaylists | PASS | 报告 |
| 92 | NotesIsTodo | PASS | 报告 |
| 93 | NotesMeetingAttendeeCount | PASS | 报告 |
| 94 | NotesRecipeIngredientCount | PASS | 报告 |
| 95 | NotesTodoItemCount | PASS | 报告 |
| 96 | SimpleCalendarAnyEventsOnDate | PASS | 报告 |
| 97 | SimpleCalendarEventOnDateAtTime | PASS | 报告 |
| 98 | SimpleCalendarEventsInNextWeek | PASS | 报告 |
| 99 | SimpleCalendarEventsInTimeRange | PASS | 报告 |
| 100 | SimpleCalendarEventsOnDate | PASS | 报告 |
| 101 | SimpleCalendarFirstEventAfterStartTime | PASS | 报告 |
| 102 | SimpleCalendarLocationOfEvent | PASS | 报告 |
| 103 | SimpleCalendarNextEvent | PASS | 报告 |
| 104 | SimpleCalendarNextMeetingWithPerson | PASS | 报告 |
| 105 | SportsTrackerActivitiesCountForWeek | PASS | 报告 |
| 106 | SportsTrackerActivitiesOnDate | PASS | 报告 |
| 107 | SportsTrackerActivityDuration | PASS | 报告 |
| 108 | SportsTrackerLongestDistanceActivity | PASS | 报告 |
| 109 | SportsTrackerTotalDistanceForCategoryOverInterval | PASS | 报告 |
| 110 | SportsTrackerTotalDurationForCategoryThisWeek | PASS | 报告 |
| 111 | TasksCompletedTasksForDate | PASS | 报告 |
| 112 | TasksDueNextWeek | PASS | 报告 |
| 113 | TasksDueOnDate | PASS | 报告 |
| 114 | TasksHighPriorityTasks | PASS | 报告 |
| 115 | TasksHighPriorityTasksDueOnDate | PASS | 报告 |
| 116 | TasksIncompleteTasksOnDate | PASS | 报告 |
Round 2（7 份报告 · 3 PASS · 4 FAIL）
| # | 任务 | 状态 | 报告 |
| -- | --------------------------------- | -------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 37 | MarkorTranscribeVideo | FAIL | 报告 |
| 40 | OsmAndMarker | FAIL | 报告 |
| 41 | OsmAndTrack | PASS | 报告 |
| 43 | RecipeAddMultipleRecipesFromImage | PASS | 报告 |
| 48 | RecipeDeleteDuplicateRecipes2 | FAIL | 报告 |
| 49 | RecipeDeleteDuplicateRecipes3 | FAIL | 报告 |
| 83 | SystemCopyToClipboard | PASS | 报告 |
Round 3（5 份报告 · 2 PASS · 3 FAIL）
| # | 任务 | 状态 | 报告 |
| -- | ----------------------------- | -------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 15 | ExpenseAddMultipleFromMarkor | PASS | 报告 |
| 37 | MarkorTranscribeVideo | FAIL | 报告 |
| 40 | OsmAndMarker | PASS | 报告 |
| 48 | RecipeDeleteDuplicateRecipes2 | FAIL | 报告 |
| 49 | RecipeDeleteDuplicateRecipes3 | FAIL | 报告 |