Same Tasks, Different Apps: Why Mobile GUI Agents Fail to Generalize?

Tien Tran; Namho Koh; Daiki E. Matsunaga; Ayush Jain; Kee Eung Kim

cs.AISep 28, 2026

Same Tasks, Different Apps: Why Mobile GUI Agents Fail to Generalize?

Authors: Tien Tran, Namho Koh, Daiki E. Matsunaga, Ayush Jain, Kee Eung Kim

Organizations: KAIST · CMU

Abstract

Mobile GUI agents deployed in real settings must work across different applications that support the same functionality. Most existing benchmarks test each task in only one app, so a high score can mean the agent understands the task, or only that it knows that particular app. We introduce AnyAppBench, a category-controlled live Android benchmark that evaluates cross-application generalization while keeping the user goal fixed. It spans 10 functional categories, 100 task templates, and 520 task--application pairs over 52 applications. Agents run from raw instructions and with app-independent sub-goals, and a VLM judge labels every failed run under a fixed failure taxonomy whose reliability is measured by human annotation. We find that, across 13 agents, success on the original application does not transfer reliably to new applications with the same goal. Furthermore, providing high-level sub-goal decomposition produces only small, category-dependent changes that do not close the gap, and the mix of failure types changes with the target interface. Based on those insights, we believe the AnyAppBench benchmark provides an important stepping stone toward robust real-world deployment of mobile GUI agents. Our code, data and the leaderboard can be found at the project website https://anyappbench.github.io/.

Figures & tables

Figure 1: Overview of AnyAppBench. Single-application evaluation conflates task difficulty, agent strategy, and interface effects. AnyAppBench fixes the user goal and semantic success condition while varying the application within a functional category. Each task–application pair runs on live Android under direct (C1) and application-independent sub-goal-augmented (C2) conditions. Deterministic programmatic verifiers alone decide success. A VLM judge labels every verifier-failed C1 trajectory under a fixed failure taxonomy, and a second judge plus human annotation measure how reliable those labels are. The benchmark spans 10 categories, 100 task templates, 52 applications, and 520 validated task–application pairs.

Figure 2: Construction of AnyAppBench. We organize app-independent task templates into 10 functional categories and pair each category’s 10 templates only with applications from the same category, producing 520 task–application pairs across 52 applications. A pair is retained only if it preserves the user-level goal, admits an equivalent programmatic verifier, runs end-to-end in the common harness, and is human-solvable using the agent’s primitive action space.

	All	AW app	New	Gap
C1 (instruction)	37.8	49.4	34.5	14.9
C2 ( $+$ sub-goals)	40.9	55.6	36.8	18.8
$\Delta$ (C2 $-$ C1)	+3.2 ∗	+6.2	+2.3	+3.9
95% CI	+0.3, +6.1	$-$ 2.0, +14.8	$-$ 0.6, +5.2	$-$ 4.4, +13.1

Table 2: Effect of app-independent sub-goal augmentation (2,285 matched pairs over the two seeded runs). C1 uses the original instruction. C2 prepends Gemini-3.1-Pro-generated app-independent sub-goals. Gap

=

success on the original AndroidWorld apps minus success on the new apps. The gap widens under C2; the interaction is a direction, not a precise size (see the appendix).

Figure 3: Six-class automated failure attribution. Per-agent distributions assigned by Gemini 3.1 Pro (G) and Qwen3-VL-30B-A3B (Q) on the same verifier-failed C1 episodes. Bars show the share assigned to each failure class; $n$ reports the number of matched episodes for each agent.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Live interaction	Fixed goal across alternative apps	Within-category app control	Trajectory-level failure diagnosis
AndroidWorld ( Rawles et al., 2025 )	✓	✗	✗	✗
AndroidLab ( Xu et al., 2025b )	✓	✗	✗	✗
SPA-Bench ( Chen et al., 2025a )	✓	✗ $\dagger$	✗	✗
GUI-Odyssey ( Lu et al., 2025a )	✗	✗ $\dagger$	✗	✗
TransBench ( Lu et al., 2025b )	✗	Partial $\ddagger$	✗	✗
AnyAppBench (ours)	✓	✓	✓	✓

Appendix

Table A.1: Comparison with representative mobile-GUI benchmarks. AnyAppBench is the only benchmark shown that satisfies all four criteria. It contains 520 validated task–app pairs, and a VLM judge diagnoses every failed trajectory, with human annotations validating the judge (Appendix C ).

Category	Task templates and verifier evidence type
Tasks / To-Do	CompletedTasksForDate AW ( U ), DueNextWeek AW ( U ), DueOnDate AW ( U ), HighPriorityTasks AW ( U ), HighPriorityDueOnDate AW ( U ), IncompleteTasksOnDate AW ( U ), DueWithTime ( D ), Recurring ( D ), EditTask ( D ), CompleteTask ( D )
Notes	CreateNote AW ( D ), EditNote AW ( D ), MergeNotes AW ( D ), DeleteNote AW ( D ), SearchNote AW ( U ), ShareOrImportNote AW ( H ), FolderOrMoveNote AW ( D ), AttachOrTranscribeContent AW ( H ), CreateChecklist AW ( D ), CountTodoItems ( U )
Finance	AddExpense AW ( D ), AddMultipleExpenses AW ( D ), DeleteTransaction AW ( D ), EditTransaction AW ( D ), AddIncome AW ( D ), DeleteDuplicateTransactions ( D ), AttachReceipt ( H ), CategorySummary ( U ), DateRangeTotal ( U ), TransferBetweenWallets ( D )
Music	CreatePlaylist AW ( D ), AddToPlaylist AW ( D ), AddToQueue AW ( U ), RemoveFromPlaylist AW ( D ), RenamePlaylist AW ( D ), SaveOrExportPlaylist ( H ), PlaylistDuration ( U ), ReorderQueue ( U ), SleepTimer ( U ), SearchAndPlay ( U )
Calendar	AddOneEvent AW ( D ), AddTimedEvent AW ( D ), AddRepeatingEvent AW ( D ), DeleteEvent AW ( D ), EditEvent AW ( D ), MoveEvent AW ( D ), EventsOnDate AW ( U ), NextEvent ( U ), EventsInRange ( U ), AddReminder ( D )
SMS	Send AW ( D ), Reply AW ( D ), ReplyMostRecent AW ( D ), Resend AW ( D ), SendClipboard AW ( D ), SendReceivedAddress AW ( D ), CreateDraftMessage ( D ), EditDraftMessage ( D ), DeleteConversation ( D ), ForwardMessage ( D )

Appendix

Table B.1: Verifier evidence type by task template. D : durable-state verifier. U : UI-state verifier. H : verifier combining durable and UI evidence or whose evidence source depends on the application. Templates marked AW preserve AndroidWorld task intent (48 of 100).

Partition	Categories	Selection	Episodes	Application role
First five	Tasks, Notes, Finance, Music, Calendar	Fixed-seed uniform	333	68 original-app / 265 new-app
Remaining five	SMS, File Manager, Maps, Contacts, Clock	Fixed-seed, $\geq$ 10% per agent	226 of 519	32 original-app / 194 new-app
Full reference	All ten	Combined	559	100 original-app / 459 new-app

Appendix

Table C.1: Human-annotation sampling partitions. First-partition counts are taken from its frozen annotation ledger, and second-partition counts are taken from the archived submission freeze of the annotation file.

Failure label	First five ( $N=333$ )	Remaining five ( $N=226$ )	All ( $N=559$ )
Planning	50	101	151
Grounding	129	65	194
Mixed	27	24	51
Execution/tooling	76	10	86
Environment/evaluator	45	17	62
Unknown	6	9	15

Appendix

Table C.2: Human failure-label counts. Counts describe the annotated sample, not population failure rates. The combined column concatenates the two partitions’ episode-level labels.

Figure 4: Human annotation interface. Web-based tool used by the annotator to review trajectory evidence and assign one diagnostic failure label per episode while all automated-judge outputs are hidden.

Figure 5: Programmatic success context for failure attribution. Equal-app-weighted C1 success on AndroidWorld-original and new applications, reported by agent (a) and category (b) for SMS, File Manager, Maps, Contacts, and Clock. Programmatic verifiers determine these outcomes. The VLM judges are applied only after verification to diagnose failed episodes.

Judge	Configuration
Gemini 3.1 Pro ( gemini-3.1-pro-preview )	Two-stage progressive judge (triage $\rightarrow$ deep review); primary automated sensitivity judge
Qwen3-VL-30B-A3B-Instruct (vLLM, 32k context)	Single-call re-judge of the same cases; independent judge-model sensitivity check

Appendix

Table D.1: Automated judges used for sensitivity analysis.

		Tasks.org		Cfait		Todo List (PFA)		ntodotxt		TaskMate		Grit
Family	Model	AW	New	AW	New	AW	New	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	27.8	8.3	38.9	8.3	22.2	8.3	22.2	0.0	22.2	0.0	22.2	0.0
	UI-Venus-7B	77.8	33.3	22.2	0.0	22.2	8.3	38.9	0.0	55.6	33.3	22.2	8.3
	GUI-Owl-7B	61.1	33.3	22.2	0.0	22.2	8.3	22.2	0.0	55.6	33.3	22.2	8.3
	MAI-UI-8B	94.4	75.0	72.2	0.0	22.2	8.3	22.2	8.3	55.6	8.3	88.9	33.3
	UI Voyager-4B	77.8	0.0	38.9	8.3	55.6	8.3	88.9	0.0	72.2	0.0	88.9	8.3

Appendix

Table E.1: App-level success rates for Tasks / To-Do (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (6) and newly added AnyAppBench templates (4), with three runs per template. Each cell averages 18 or 12 rollouts.

Family	Model	Joplin		Markor		NotallyX		neutriNote CE		Notesnook		Orgzly Revived
Family	Model	AW	New	AW	New	AW	New	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	29.6	0.0	0.0	0.0	11.1	0.0	3.7	33.3	7.4	0.0	7.4	0.0
	UI-Venus-7B	40.7	0.0	11.1	0.0	29.6	0.0	18.5	0.0	40.7	33.3	40.7	0.0
	GUI-Owl-7B	40.7	0.0	3.7	0.0	59.3	33.3	40.7	33.3	18.5	33.3	44.4	33.3
	MAI-UI-8B	77.8	66.7	59.3	0.0	51.9	33.3	3.7	33.3	51.9	0.0	44.4	0.0
	UI Voyager-4B	25.9	0.0	3.7	0.0	59.3	0.0	51.9	0.0	11.1	0.0	33.3	33.3

Appendix

Table E.2: App-level success rates for Notes (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (9) and newly added AnyAppBench templates (1), with three runs per template. Each cell averages 27 or 3 rollouts.

		Pro Expense		Oinkoin		OpenMoney Box		My Expenses		Finance Manager		Sushi
Family	Model	AW	New	AW	New	AW	New	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	33.3	0.0	20.0	13.3	6.7	0.0	26.7	0.0	6.7	6.7	6.7	0.0
	UI-Venus-7B	33.3	0.0	26.7	6.7	26.7	0.0	26.7	20.0	6.7	6.7	46.7	6.7
	GUI-Owl-7B	13.3	0.0	20.0	13.3	26.7	6.7	6.7	6.7	20.0	13.3	20.0	13.3
	MAI-UI-8B	93.3	13.3	60.0	13.3	46.7	6.7	53.3	6.7	66.7	0.0	60.0	13.3
	UI Voyager-4B	93.3	33.3	6.7	6.7	46.7	0.0	53.3	0.0	46.7	6.7	26.7	0.0

Appendix

Table E.3: App-level success rates for Finance (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (5) and newly added AnyAppBench templates (5), with three runs per template. Each cell averages 15 rollouts.

		Retro Music		Fossify Music Player		Apollo		SicMu Neo		Phonograph Plus		Monster Music
Family	Model	AW	New	AW	New	AW	New	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	46.7	20.0	13.3	6.7	13.3	0.0	20.0	0.0	20.0	0.0	40.0	0.0
	UI-Venus-7B	26.7	20.0	13.3	6.7	13.3	0.0	20.0	0.0	20.0	0.0	40.0	0.0
	GUI-Owl-7B	26.7	13.3	13.3	0.0	13.3	0.0	20.0	0.0	20.0	0.0	20.0	0.0
	MAI-UI-8B	26.7	13.3	33.3	0.0	20.0	0.0	20.0	6.7	20.0	0.0	40.0	6.7
	UI Voyager-4B	26.7	20.0	13.3	0.0	20.0	0.0	20.0	6.7	20.0	0.0	20.0	0.0

Appendix

Table E.4: App-level success rates for Music (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (5) and newly added AnyAppBench templates (5), with three runs per template. Each cell averages 15 rollouts.

		Simple Calendar Pro		Etar		Fossify Calendar		Calendar		KashCal
Family	Model	AW	New	AW	New	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	14.3	0.0	0.0	0.0	14.3	11.1	0.0	0.0	9.5	11.1
	UI-Venus-7B	61.9	44.4	9.5	11.1	42.9	11.1	0.0	0.0	23.8	0.0
	GUI-Owl-7B	4.8	0.0	0.0	0.0	14.3	0.0	0.0	0.0	0.0	0.0
	MAI-UI-8B	4.8	0.0	9.5	0.0	71.4	77.8	23.8	0.0	38.1	11.1
	UI Voyager-4B	33.3	44.4	9.5	0.0	28.6	11.1	28.6	0.0	23.8	11.1

Appendix

Table E.5: App-level success rates for Calendar (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (7) and newly added AnyAppBench templates (3), with three runs per template. Each cell averages 21 or 9 rollouts.

		Simple SMS Messenger		Fossify Messages		QUIK SMS		Messages
Family	Model	AW	New	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	72.2	75.0	77.8	75.0	66.7	58.3	66.7	66.7
	UI-Venus-7B	27.8	25.0	38.9	8.3	22.2	25.0	44.4	16.7
	GUI-Owl-7B	72.2	58.3	66.7	50.0	38.9	25.0	77.8	50.0
	MAI-UI-8B	100.0	66.7	72.2	58.3	88.9	66.7	83.3	83.3
	UI Voyager-4B	94.4	25.0	77.8	16.7	61.1	25.0	72.2	8.3

Appendix

Table E.6: App-level success rates for SMS (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (6) and newly added AnyAppBench templates (4), with three runs per template. Each cell averages 18 or 12 rollouts.

		Material Files		Amaze File Manager		Fossify File Manager		Total Commander		X-plore File Manager
Family	Model	AW	New	AW	New	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	50.0	54.2	66.7	54.2	50.0	58.3	50.0	25.0	33.3	16.7
	UI-Venus-7B	50.0	16.7	33.3	54.2	50.0	16.7	0.0	0.0	0.0	12.5
	GUI-Owl-7B	50.0	37.5	16.7	25.0	0.0	25.0	16.7	0.0	16.7	16.7
	MAI-UI-8B	83.3	58.3	100.0	66.7	100.0	70.8	83.3	41.7	33.3	54.2
	UI Voyager-4B	50.0	29.2	50.0	33.3	50.0	25.0	50.0	0.0	16.7	20.8

Appendix

Table E.7: App-level success rates for File Manager (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (2) and newly added AnyAppBench templates (8), with three runs per template. Each cell averages 6 or 24 rollouts.

		OsmAnd		Organic Maps		CoMaps
Family	Model	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	0.0	0.0	22.2	33.3	0.0	28.6
	UI-Venus-7B	0.0	4.8	44.4	33.3	11.1	42.9
	GUI-Owl-7B	44.4	23.8	22.2	33.3	11.1	23.8
	MAI-UI-8B	44.4	23.8	22.2	42.9	0.0	28.6
	UI Voyager-4B	44.4	38.1	44.4	42.9	44.4	28.6

Appendix

Table E.8: App-level success rates for Maps (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (3) and newly added AnyAppBench templates (7), with three runs per template. Each cell averages 9 or 21 rollouts.

		Google Contacts		Fossify Contacts		Connect You		Simple Contacts Pro SE		Right Contact
Family	Model	AW	New	AW	New	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	100.0	66.7	50.0	4.2	0.0	0.0	100.0	70.8	0.0	8.3
	UI-Venus-7B	50.0	12.5	50.0	0.0	0.0	0.0	100.0	25.0	0.0	0.0
	GUI-Owl-7B	100.0	41.7	0.0	0.0	0.0	0.0	100.0	54.2	50.0	0.0
	MAI-UI-8B	100.0	79.2	50.0	29.2	0.0	0.0	16.7	33.3	16.7	0.0
	UI Voyager-4B	100.0	62.5	0.0	0.0	0.0	4.2	0.0	0.0	0.0	0.0

Appendix

Table E.9: App-level success rates for Contacts (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (2) and newly added AnyAppBench templates (8), with three runs per template. Each cell averages 6 or 24 rollouts.

		Google Clock		Clock		Simple Clock		Clock You		Chrono		Fossify Clock
Family	Model	AW	New	AW	New	AW	New	AW	New	AW	New	AW	New
GUI Trained	UI-Venus-Navi-72B	66.7	81.0	0.0	14.3	66.7	71.4	66.7	14.3	33.3	57.1	66.7	71.4
	UI-Venus-7B	66.7	90.5	0.0	0.0	66.7	47.6	0.0	0.0	66.7	57.1	77.8	47.6
	GUI-Owl-7B	66.7	76.2	22.2	4.8	66.7	52.4	66.7	23.8	66.7	61.9	66.7	52.4
	MAI-UI-8B	66.7	95.2	66.7	76.2	44.4	85.7	66.7	19.0	66.7	95.2	55.6	66.7
	UI Voyager-4B	88.9	81.0	11.1	19.0	66.7	28.6	66.7	19.0	66.7	47.6	66.7	28.6

Appendix

Table E.10: App-level success rates for Clock (%, three runs pooled). Each app has two sub-columns: AW-inherited templates (3) and newly added AnyAppBench templates (7), with three runs per template. Each cell averages 9 or 21 rollouts.

Category	Task Templates	AW Orig. App(s)	Newly Installed Apps	# Apps AW+New
Tasks / To-Do	CompletedTasksForDate AW , DueNextWeek AW , DueOnDate AW , HighPriorityTasks AW , HighPriorityDueOnDate AW , IncompleteTasksOnDate AW , DueWithTime, Recurring, EditTask, CompleteTask	Tasks.org	Cfait, Todo List (PFA), ntodotxt, TaskMate, Grit	1+ 5
Notes	CreateNote AW , EditNote AW , MergeNotes AW , DeleteNote AW , SearchNote AW , ShareOrImportNote AW , FolderOrMoveNote AW , AttachOrTranscribeContent AW , CountTodoItems AW , CreateChecklist	Joplin , Markor	NotallyX, neutriNote CE, Notesnook, Orgzly Revived	2+ 4
Finance	AddExpense AW , AddMultipleExpenses AW , DeleteTransaction AW , DeleteDuplicateTransactions AW , AttachReceipt AW , AddIncome, EditTransaction, CategorySummary, DateRangeTotal, TransferBetweenWallets	Pro Expense	Oinkoin, OpenMoneyBox, My Expenses, Finance Manager, Sushi	1+ 5
Music	CreatePlaylist AW , AddToPlaylist AW , AddToQueue AW , SaveOrExportPlaylist AW , PlaylistDuration AW , RenamePlaylist, RemoveFromPlaylist, ReorderQueue, SleepTimer, SearchAndPlay	Retro Music	Fossify Music Player, Apollo, SicMu Neo, Phonograph Plus, MonsterMusic	1+ 5
Calendar	AddOneEvent AW , AddTimedEvent AW , AddRepeatingEvent AW , DeleteEvent AW , EventsOnDate AW , NextEvent AW , EventsInRange AW , EditEvent, AddReminder, MoveEvent	Simple Calendar Pro	Etar, Fossify Calendar, Calendar, KashCal	1+ 4
SMS	Send AW , Reply AW , ReplyMostRecent AW , Resend AW , SendClipboard AW , SendReceivedAddress AW , CreateDraftMessage, EditDraftMessage, DeleteConversation, ForwardMessage	Simple SMS Messenger	Fossify Messages, QUIK SMS, Messages	1+ 3

Appendix

Table E.11: AnyAppBench task and application inventory. Each of the 10 categories contains 10 task templates. AW marks AndroidWorld-derived task intents, bold app names mark AndroidWorld-original canonical applications, and the remaining apps are alternatives. The inventory contains 48 AndroidWorld-derived and 52 newly added templates across 11 canonical and 41 alternative applications, yielding 520 category-matched task–application pairs.

Figure 6: Same task, different application. GPT-5.1 performs the same semantic task (create a contact with a given phone number and mark it as a favorite) on two applications of the Contacts category. Top: on Google Contacts, the AndroidWorld-original app, the agent completes all fields, saves, and stars the contact (green blocks, verifier success). Bottom: on Right Contact, a new app, the agent enters the name and number but never completes the save and favorite flow (red blocks, verifier failure). The final frame of each strip carries the verifier outcome.

Explore similar work

APPSim-Bench: Bridging Real-world Apps and Reproducible Evaluation for Mobile GUI Agents
Sep 7, 2026Jintian Feng, Long Chen, Xiao Yu +7Graphical User Interface Agents Device
Laboratory for Artificial Intelligence and New Forms of Education, Central China Normal University · Faculty of Artificial Intelligence in Education, Central China Normal University · Agentic Labs, Acrab AI
JarvisGUI: Towards Cross-Device GUI Agents with Dynamic Task Composition
Sep 9, 2026Zixiang Chen, Yuheng Lu, Zihao Cheng +8Graphical User Interface Agents Graphical User Interface
School of Computer Science and Engineering, Beihang University, Beijing, China · Beijing Institute of Technology, Beijing, China · Baidu Inc., Beijing, China
AndroidDaily: A Verifiable Benchmark for Mobile GUI Agents on Real-World Closed-Source Applications
May 26, 2026Yifan Sui, Xin Huang, Hongbing Li +14Device Graphical User Interface Agents
1Beijing University of Posts and Telecommunications · 2StepFun · 3Waseda University

Sep 7, 2026, cs.AIJ/K move · Enter open · S save

APPSim-Bench: Bridging Real-world Apps and Reproducible Evaluation for Mobile GUI Agents

Jintian Feng, Long Chen, Xiao Yu, Jiayi Dai+6

Laboratory for Artificial Intelligence and New Forms of Education, Central China Normal University · Faculty of Artificial Intelligence in Education, Central China Normal University · Agentic Labs, Acrab AI

Graphical User Interface Agents Device Evaluation Benchmarks Reproducibility

Open paper AskPDF arXiv

Mobile GUI agents can execute tasks from natural-language instructions, but their evaluation remains difficult to make both realistic and reproducible. Existing benchmarks typically trade off these goals: simplified apps lack real-world mobile complexity, whereas live commercial apps introduce uncontrolled variation from recommendations, advertisements, accounts, and changing content. We propose AppSim-Bench, which addresses this trade-off through controllable simulated apps that preserve task-relevant interaction logic while supporting deterministic evaluation. Built through a coding-agent-assisted and human-verified workflow, it contains 557 tasks across 17 high-frequency Chinese and English apps. Its controllable backend data and outcome-based verification remove major sources of environmental stochasticity, enabling reproducible cross-model comparison. Evaluating 19 GUI agents, spanning general-purpose and GUI-specialized systems, we find that autonomous mobile execution remains far from solved. The best model completes only 50.27% of tasks, and 28.55% of tasks are not solved by any agent. Further analysis shows that failures concentrate in longer workflows, numerical reasoning tasks, and inefficient trajectories marked by high action overhead and budget exhaustion. Our project is available at https://github.com/Acrab-Agentic-Labs/AppSim.

Abstract

Explore similar work

APPSim-Bench: Bridging Real-world Apps and Reproducible Evaluation for Mobile GUI Agents

JarvisGUI: Towards Cross-Device GUI Agents with Dynamic Task Composition

AndroidDaily: A Verifiable Benchmark for Mobile GUI Agents on Real-World Closed-Source Applications