pith. sign in

arxiv: 2509.06477 · v2 · submitted 2025-09-08 · 💻 cs.AI

MAS-Bench: A Unified Benchmark for Shortcut-Augmented Hybrid Mobile GUI Agents

Pengxiang Zhao , Guangyi Liu , YaoZhen Liang , Weiqing He , Zhengxi Lu , Wenhao Wang , Yuehao Huang , Yuxiang Chai
show 6 more authors
Zhaolu Kang Yaxuan Guo Hao Wang Kexin Zhang Liang Liu Yong Liu
This is my paper

Pith reviewed 2026-05-18 18:38 UTC · model grok-4.3

classification 💻 cs.AI
keywords mobile GUI agentshybrid agentsshortcutsbenchmarkdeep linksAPIsmobile automationMLLM agents
0
0 comments X

The pith

A benchmark for hybrid mobile agents shows they reach 68.3 percent success and 39 percent higher efficiency by combining GUI actions with shortcuts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces MAS-Bench to test agents that mix flexible screen-based GUI operations with faster shortcuts such as APIs and deep links for mobile tasks. It evaluates both the use of a fixed set of shortcuts and an agent's ability to discover and build new reusable ones on its own. Tests across 139 tasks in 11 apps show clear gains in success rate and speed compared with agents that rely only on GUI steps. The same setup also measures how well agent-created shortcuts perform relative to the predefined ones.

Core claim

MAS-Bench supplies 139 complex tasks spanning 11 real-world applications, a knowledge base of 88 predefined shortcuts consisting of APIs, deep-links, and RPA scripts, and nine metrics to score hybrid agents. Hybrid agents that employ both GUI operations and shortcuts achieve success rates up to 68.3 percent while completing tasks 39 percent more efficiently than GUI-only agents. The evaluation framework further exposes measurable quality differences between the predefined shortcuts and those generated autonomously by the agents.

What carries the argument

MAS-Bench benchmark that supplies tasks, a shortcut knowledge base, and metrics to measure both execution of hybrid mobile agents and their ability to generate new shortcuts.

If this is right

  • Hybrid agents that combine GUI steps with shortcuts can be deployed to complete more mobile tasks reliably and with less time.
  • Agents should include mechanisms to detect reusable low-cost workflows and store them as new shortcuts.
  • Developers can now compare different shortcut-generation methods using the same set of tasks and metrics.
  • Training pipelines for mobile agents can incorporate both predefined and self-generated shortcuts to improve overall performance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the benchmark tasks generalize well, the same evaluation approach could be adapted to measure hybrid agents on desktop or web environments.
  • Closing the quality gap between predefined and agent-generated shortcuts might require new algorithms that explicitly optimize for reusability across multiple tasks.
  • Embedding shortcut generation directly into the agent's reasoning loop could reduce the need for large external knowledge bases over time.

Load-bearing premise

The 139 tasks across 11 applications capture enough of real-world mobile automation that results on this benchmark will predict how agents perform on unseen apps and tasks.

What would settle it

Test the same hybrid agents on a fresh collection of mobile applications and tasks outside the original 11 apps and check whether the reported success-rate and efficiency advantages remain intact.

Figures

Figures reproduced from arXiv: 2509.06477 by Guangyi Liu, Hao Wang, Kexin Zhang, Liang Liu, Pengxiang Zhao, Weiqing He, Wenhao Wang, YaoZhen Liang, Yaxuan Guo, Yong Liu, Yuehao Huang, Yuxiang Chai, Zhaolu Kang, Zhengxi Lu.

Figure 1
Figure 1. Figure 1: Workflow of GUI Only vs. GUI-Shortcut Hy￾brid Agent. Shortcuts improve agent execution efficiency by bypassing GUI operations. As GUI agents are increasingly applied to complex tasks, enhancing their efficiency has become a central research focus. In this context, hybrid paradigms have emerged as a promising solution, demonstrating effectiveness across a wide range of platforms (Wang et al. 2025b; Zheng et… view at source ↗
Figure 2
Figure 2. Figure 2: Functional Comparison of APIs, Deep-Links, and RPA Scripts. The figure uses the Amazon app as an ex￾ample: (a) the open cart() API call directly opens the shop￾ping cart; (b) the search product() deep-link directly per￾forms a product search; and (c) an RPA script combines APIs, deep links, and GUI operations to automate a com￾plete workflow. 3.3 Shortcut Knowledge Base Predefined Shortcuts Knowledge Base.… view at source ↗
Figure 3
Figure 3. Figure 3: Evaluation Workflow for Agents’ Shortcut Gen￾eration Capability. The process consists of two stages: (a) Shortcut Generation Stage, where the agent creates its short￾cut knowledge base; and (b) Quality Evaluation Stage. The Agent-generated shortcuts is imported into a baseline agent for performance testing. The quality of the Agent-generated shortcuts is then measured by comparing this performance against … view at source ↗
Figure 4
Figure 4. Figure 4: The pipeline of MAS-Bench. The GUI-Shortcut agent first filters products using the search product shortcut, selects an item via GUI operations, and then adds it to the cart using the add to cart shortcut. The entire process is monitored by an automated evaluation module, which outputs metrics such as success rate and efficiency. number of steps required to complete the task. A lower ratio indicates better … view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison of MAS-MobileAgent with and without shortcuts. The base models are Gemini￾2.5-Pro and Gemini-2.0-Flash. Data points show the rela￾tionship between SR and MET for single-app and cross￾app tasks, with circle size representing mean cost. Results demonstrate that shortcuts benefit both models, with more significant improvements for the weaker Gemini-2.0-Flash. Step Ratio on Successful ta… view at source ↗
Figure 6
Figure 6. Figure 6: As detailed in Table 8, this process resulted in 46 [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Examples of GUI-shortcut hybrid agent failure cases. (a) Selection and Planning Error: The agent incorrectly invokes [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
read the original abstract

Shortcuts such as APIs and deep-links have emerged as efficient complements to flexible GUI operations, fostering a promising hybrid paradigm for MLLM-based mobile automation. However, systematic evaluation of GUI-shortcut hybrid agents remains largely underexplored. To bridge this gap, we introduce MAS-Bench, a benchmark that pioneers the evaluation of GUI-shortcut hybrid agents with a specific focus on the mobile domain. Beyond merely using predefined shortcuts, MAS-Bench assesses an agent's capability to autonomously generate shortcuts by discovering and creating reusable, low-cost workflows. It features 139 complex tasks across 11 real-world applications, a knowledge base of 88 predefined shortcuts (APIs, deep-links, RPA scripts), and 9 evaluation metrics. Experiments demonstrate that hybrid agents achieve up to 68.3% success rate and 39% greater execution efficiency than GUI-only counterparts. Furthermore, our evaluation framework effectively reveals the quality gap between predefined and agent-generated shortcuts, validating its capability to assess shortcut generation methods. MAS-Bench addresses the lack of systematic benchmarks for GUI-shortcut hybrid mobile agents, providing a foundational platform for future advancements in creating more efficient and robust intelligent agents. Project page: https://pengxiang-zhao.github.io/MAS-Bench.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces MAS-Bench, a benchmark for systematic evaluation of GUI-shortcut hybrid agents in the mobile domain. It comprises 139 complex tasks across 11 real-world applications, a knowledge base of 88 predefined shortcuts (APIs, deep-links, RPA scripts), and 9 evaluation metrics. Experiments report that hybrid agents reach up to 68.3% success rate and 39% greater execution efficiency than GUI-only baselines, while the framework detects quality gaps between predefined and agent-generated shortcuts. The work positions MAS-Bench as a foundational platform to advance research on efficient MLLM-based mobile automation.

Significance. If the task set is representative, the benchmark fills a clear gap by enabling controlled comparison of hybrid versus GUI-only agents and by providing a mechanism to assess autonomous shortcut generation. The concrete empirical deltas on success rate and efficiency supply a reproducible starting point for future work; the emphasis on both predefined and generated shortcuts is a useful distinction not commonly isolated in prior GUI-agent evaluations.

major comments (1)
  1. [§3 and §4] §3 (Benchmark Construction) and §4 (Experiments): No selection criteria, diversity metrics, category coverage statistics, or cross-app generalization analysis are provided for the 11 applications or 139 tasks. Because the headline claims (68.3% success, 39% efficiency gain, and the assertion that MAS-Bench is a 'foundational platform') rest on the assumption that performance on this set predicts behavior on unseen apps and tasks, the absence of such justification is load-bearing and must be addressed before the generalization statements can be accepted.
minor comments (2)
  1. [Abstract] Abstract: The nine evaluation metrics are referenced but neither named nor briefly characterized; a short enumeration would improve readability for readers who do not immediately consult the full text.
  2. [§4] Notation: The distinction between 'predefined' and 'agent-generated' shortcuts is central yet occasionally blurred in the experimental narrative; consistent terminology and a clear table mapping each metric to the two categories would reduce ambiguity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We are grateful to the referee for highlighting an important aspect of our benchmark construction. The feedback points to a need for greater transparency in how the tasks and applications were selected, which is crucial for validating the broader applicability of our results. We respond to this comment below and will make corresponding revisions to the manuscript.

read point-by-point responses

  1. Referee: [§3 and §4] §3 (Benchmark Construction) and §4 (Experiments): No selection criteria, diversity metrics, category coverage statistics, or cross-app generalization analysis are provided for the 11 applications or 139 tasks. Because the headline claims (68.3% success, 39% efficiency gain, and the assertion that MAS-Bench is a 'foundational platform') rest on the assumption that performance on this set predicts behavior on unseen apps and tasks, the absence of such justification is load-bearing and must be addressed before the generalization statements can be accepted.

    Authors: We fully agree with the referee that explicit selection criteria and diversity analysis are necessary to substantiate the generalization claims. The original submission omitted a detailed description of how the 11 applications were chosen and the distribution of the 139 tasks. In the revised version, we will expand §3 to include: (1) Selection criteria: Applications were chosen based on popularity in app stores and to represent diverse functionalities and UI complexities in the mobile domain. Tasks were designed as complex, multi-step operations that could benefit from shortcut augmentation. (2) Diversity metrics: We will add statistics such as the number of tasks per application, distribution across different app types, and average task complexity. (3) A discussion on cross-app generalization, acknowledging that while we have not performed exhaustive experiments on completely unseen applications, the selected set aims to be representative, and we will include this as a limitation and direction for future work. We will also adjust the language in the abstract and introduction to present MAS-Bench as a foundational benchmark for this emerging area while being clear about the scope of the current evaluation. These changes will strengthen the manuscript and address the concerns raised. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark results are direct measurements

full rationale

The paper introduces MAS-Bench as a new benchmark consisting of 139 tasks across 11 apps, 88 predefined shortcuts, and 9 metrics, then reports direct experimental outcomes such as hybrid agents reaching 68.3% success rate and 39% greater efficiency than GUI-only baselines. These figures are obtained by running agents on the benchmark tasks and measuring success and efficiency; no equations, fitted parameters, predictions, or first-principles derivations are presented that could reduce to the inputs by construction. The evaluation of quality gaps between predefined and agent-generated shortcuts is likewise an observational comparison on the held-out tasks rather than a self-referential loop. The central claims therefore remain independent and falsifiable through replication on the released benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The benchmark rests on the domain assumption that the chosen 11 apps and 139 tasks form a sufficient test distribution for hybrid agents; no free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption The selected mobile applications and tasks are representative of real-world automation scenarios.
    Invoked when claiming that results on MAS-Bench will inform practical hybrid-agent development.

pith-pipeline@v0.9.0 · 5795 in / 1206 out tokens · 26509 ms · 2026-05-18T18:38:28.599211+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. OS-SPEAR: A Toolkit for the Safety, Performance,Efficiency, and Robustness Analysis of OS Agents

    cs.CL 2026-04 unverdicted novelty 7.0

    OS-SPEAR is a new evaluation toolkit that tests 22 OS agents and identifies trade-offs between efficiency and safety or robustness.

  2. FedGUI: Benchmarking Federated GUI Agents across Heterogeneous Platforms, Devices, and Operating Systems

    cs.MA 2026-04 unverdicted novelty 7.0

    FedGUI is the first comprehensive benchmark for federated GUI agents that studies cross-platform, cross-device, cross-OS, and cross-source heterogeneity, with experiments showing performance gains from cross-platform ...

  3. UI-Copilot: Advancing Long-Horizon GUI Automation via Tool-Integrated Policy Optimization

    cs.LG 2026-04 unverdicted novelty 6.0

    UI-Copilot adds a selective copilot for memory and math to GUI agents and trains tool use with separate single-turn and multi-turn optimization, yielding SOTA results on MemGUI-Bench and a 17.1% gain on AndroidWorld.

  4. Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents

    cs.LG 2026-04 unverdicted novelty 6.0

    Skill-SD turns an agent's completed trajectories into dynamic natural-language skills that condition only the teacher in self-distillation, yielding 14-42% gains over RL and OPSD baselines on multi-turn agent benchmarks.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · cited by 4 Pith papers · 8 internal anchors

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

    work page

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

    work page

  3. [3]

    Agostinelli, S.; Lupia, M.; Marrella, A.; and Mecella, M. 2022. Reactive synthesis of software robots in RPA from user interface logs. Computers in Industry, 142: 103721

    work page 2022

  4. [4]

    R.; Tramontana, P.; Ta, B

    Amalfitano, D.; Fasolino, A. R.; Tramontana, P.; Ta, B. D.; and Memon, A. M. 2014. MobiGUITAR: Automated model-based testing of mobile apps. IEEE software, 32(5): 53--59

    work page 2014

  5. [5]

    Baechler, G.; Sunkara, S.; Wang, M.; Zubach, F.; Mansoor, H.; Etter, V.; C a rbune, V.; Lin, J.; Chen, J.; and Sharma, A. 2024. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615

    work page arXiv 2024

  6. [6]

    Bridle, R.; and McCreath, E. 2006. Inducing shortcuts on a mobile phone interface. In Proceedings of the 11th international conference on Intelligent user interfaces, 327--329

    work page 2006

  7. [7]

    Chang, Y.; Wang, X.; Wang, J.; Wu, Y.; Yang, L.; Zhu, K.; Chen, H.; Yi, X.; Wang, C.; Wang, Y.; et al. 2024. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3): 1--45

    work page 2024

  8. [8]

    Chen, J.; Yuen, D.; Xie, B.; Yang, Y.; Chen, G.; Wu, Z.; Yixing, L.; Zhou, X.; Liu, W.; Wang, S.; et al. 2024. Spa-bench: A comprehensive benchmark for smartphone agent evaluation. In NeurIPS 2024 Workshop on Open-World Agents

    work page 2024

  9. [9]

    Cheng, K.; Sun, Q.; Chu, Y.; Xu, F.; Li, Y.; Zhang, J.; and Wu, Z. 2024. Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    et al., O. 2024. GPT-4 Technical Report. arXiv:2303.08774

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    Gou, B.; Wang, R.; Zheng, B.; Xie, Y.; Chang, C.; Shu, Y.; Sun, H.; and Su, Y. 2024. Navigating the Digital World as Humans Do: Universal Visual Grounding for GUI Agents. arXiv preprint arXiv:2410.05243

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    Guerreiro, T.; Gamboa, R.; and Jorge, J. 2008. Mnemonical body shortcuts: improving mobile interaction. In Proceedings of the 15th European conference on Cognitive ergonomics: the ergonomics of cool interaction, 1--8

    work page 2008

  13. [13]

    Ap- pagentx: Evolving gui agents as proficient smartphone users.arXiv preprint arXiv:2503.02268, 2025

    Jiang, W.; Zhuang, Y.; Song, C.; Yang, X.; Zhou, J. T.; and Zhang, C. 2025. AppAgentX: Evolving GUI Agents as Proficient Smartphone Users. arXiv preprint arXiv:2503.02268

    work page arXiv 2025

  14. [14]

    Kennedy, C.; and Everett, S. E. 2011. Use of cognitive shortcuts in landline and cell phone surveys. Public Opinion Quarterly, 75(2): 336--348

    work page 2011

  15. [15]

    Kirubakaran, B.; and Karthikeyani, V. 2013. Mobile application testing—Challenges and solution approach through automation. In 2013 International Conference on Pattern Recognition, Informatics and Mobile Engineering, 79--84. IEEE

    work page 2013

  16. [16]

    F.; and Klein, J

    Kong, P.; Li, L.; Gao, J.; Liu, K.; Bissyand \'e , T. F.; and Klein, J. 2018. Automated testing of android apps: A systematic literature review. IEEE Transactions on Reliability, 68(1): 45--66

    work page 2018

  17. [17]

    V.; Mayer, S.; Wei , M.; Vogelsang, J.; Weing \"a rtner, H.; and Henze, N

    Le, H. V.; Mayer, S.; Wei , M.; Vogelsang, J.; Weing \"a rtner, H.; and Henze, N. 2020. Shortcut gestures for mobile text editing on fully touch sensitive smartphones. ACM Transactions on Computer-Human Interaction (TOCHI), 27(5): 1--38

    work page 2020

  18. [18]

    Lee, J.; Lee, D.; Choi, C.; Im, Y.; Wi, J.; Heo, K.; Oh, S.; Lee, S.; and Shin, I. 2025. Safeguarding Mobile GUI Agent via Logic-based Action Verification. arXiv:2503.18492

    work page arXiv 2025

  19. [19]

    Ko, Sangeun Oh, and Insik Shin

    Lee, S.; Choi, J.; Lee, J.; Wasi, M. H.; Choi, H.; Ko, S. Y.; Oh, S.; and Shin, I. 2023. Explore, select, derive, and recall: Augmenting llm with human-like memory for mobile task automation. arXiv preprint arXiv:2312.03003

    work page arXiv 2023

  20. [20]

    Linares-V \'a squez, M.; Moran, K.; and Poshyvanyk, D. 2017. Continuous, evolutionary and large-scale: A new perspective for automated mobile app testing. In 2017 IEEE International Conference on Software Maintenance and Evolution (ICSME), 399--410. IEEE

    work page 2017

  21. [21]

    Ling, X.; Gao, M.; and Wang, D. 2020. Intelligent document processing based on RPA and machine learning. In 2020 Chinese Automation Congress (CAC), 1349--1353. IEEE

    work page 2020

  22. [22]

    Liu, G.; Zhao, P.; Liu, L.; Chen, Z.; Chai, Y.; Ren, S.; Wang, H.; He, S.; and Meng, W. 2025 a . LearnAct: Few-Shot Mobile GUI Agent with a Unified Demonstration Benchmark. arXiv preprint arXiv:2504.13805

    work page arXiv 2025

  23. [23]

    Liu, G.; Zhao, P.; Liu, L.; Guo, Y.; Xiao, H.; Lin, W.; Chai, Y.; Han, Y.; Ren, S.; Wang, H.; et al. 2025 b . Llm-powered gui agents in phone automation: Surveying progress and prospects. arXiv preprint arXiv:2504.19838

    work page arXiv 2025

  24. [24]

    Lu, Z.; Chai, Y.; Guo, Y.; Yin, X.; Liu, L.; Wang, H.; Xiong, G.; and Li, H. 2025. Ui-r1: Enhancing action prediction of gui agents by reinforcement learning. arXiv preprint arXiv:2503.21620

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  25. [25]

    Minaee, S.; Mikolov, T.; Nikzad, N.; Chenaghlu, M.; Socher, R.; Amatriain, X.; and Gao, J. 2024. Large language models: A survey. arXiv preprint arXiv:2402.06196

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    Rawles, C.; Clinckemaillie, S.; Chang, Y.; Waltz, J.; Lau, G.; Fair, M.; Li, A.; Bishop, W.; Li, W.; Campbell-Ajala, F.; et al. 2024. AndroidWorld: A dynamic benchmarking environment for autonomous agents. arXiv preprint arXiv:2405.14573

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [27]

    M.; Purohit, A

    Roffarello, A. M.; Purohit, A. K.; and Purohit, S. V. 2024. Trigger-Action Programming for Wellbeing: Insights From 6590 iOS Shortcuts. IEEE Pervasive Computing

    work page 2024

  28. [28]

    Tang, F.; Xu, H.; Zhang, H.; Chen, S.; Wu, X.; Shen, Y.; Zhang, W.; Hou, G.; Tan, Z.; Yan, Y.; et al. 2025. A Survey on (M) LLM-Based GUI Agents. arXiv preprint arXiv:2504.13865

    work page arXiv 2025

  29. [29]

    Tripathi, A. M. 2018. Learning Robotic Process Automation: Create Software robots and automate business processes with the leading RPA tool--UiPath. Packt Publishing Ltd

    work page 2018

  30. [30]

    Wang, F.; Zhang, Z.; Zhang, X.; Wu, Z.; Mo, T.; Lu, Q.; Wang, W.; Li, R.; Xu, J.; Tang, X.; et al. 2024 a . A comprehensive survey of small language models in the era of large language models: Techniques, enhancements, applications, collaboration with llms, and trustworthiness. arXiv preprint arXiv:2411.03350

    work page arXiv 2024

  31. [31]

    Wang, J.; Xu, H.; Jia, H.; Zhang, X.; Yan, M.; Shen, W.; Zhang, J.; Huang, F.; and Sang, J. 2024 b . Mobile-Agent-v2: Mobile Device Operation Assistant with Effective Navigation via Multi-Agent Collaboration. arXiv preprint arXiv:2406.01014

    work page arXiv 2024

  32. [32]

    Wang, L.; Deng, Y.; Zha, Y.; Mao, G.; Wang, Q.; Min, T.; Chen, W.; and Chen, S. 2024 c . MobileAgentBench: An Efficient and User-Friendly Benchmark for Mobile LLM Agents. arXiv preprint arXiv:2406.08184

    work page arXiv 2024

  33. [33]

    Wang, Z.; Xu, H.; Wang, J.; Zhang, X.; Yan, M.; Zhang, J.; Huang, F.; and Ji, H. 2025 a . Mobile-Agent-E: Self-Evolving Mobile Assistant for Complex Tasks. arXiv preprint arXiv:2501.11733

    work page arXiv 2025

  34. [34]

    Inducing programmatic skills for agentic tasks.arXiv preprint arXiv:2504.06821, 2025

    Wang, Z. Z.; Gandhi, A.; Neubig, G.; and Fried, D. 2025 b . Inducing programmatic skills for agentic tasks. arXiv preprint arXiv:2504.06821

    work page arXiv 2025

  35. [35]

    Wen, H.; Tian, S.; Pavlov, B.; Du, W.; Li, Y.; Chang, G.; Zhao, S.; Liu, J.; Liu, Y.; Zhang, Y.-Q.; et al. 2024. AutoDroid-V2: Boosting SLM-based GUI Agents via Code Generation. arXiv preprint arXiv:2412.18116

    work page arXiv 2024

  36. [36]

    Xu, Y.; Liu, X.; Sun, X.; Cheng, S.; Yu, H.; Lai, H.; Zhang, S.; Zhang, D.; Tang, J.; and Dong, Y. 2024. AndroidLab: Training and Systematic Benchmarking of Android Autonomous Agents. arXiv preprint arXiv:2410.24024

    work page arXiv 2024

  37. [37]

    You, K.; Zhang, H.; Schoop, E.; Weers, F.; Swearngin, A.; Nichols, J.; Yang, Y.; and Gan, Z. 2024. Ferret-UI: Grounded Mobile UI Understanding with Multimodal LLMs. arXiv preprint arXiv:2404.05719

    work page arXiv 2024

  38. [38]

    Zhang, C.; He, S.; Li, L.; Qin, S.; Kang, Y.; Lin, Q.; and Zhang, D. 2025 a . API Agents vs. GUI Agents: Divergence and Convergence. arXiv preprint arXiv:2503.11069

    work page arXiv 2025

  39. [39]

    Zhang, C.; Huang, H.; Ni, C.; Mu, J.; Qin, S.; He, S.; Wang, L.; Yang, F.; Zhao, P.; Du, C.; et al. 2025 b . Ufo2: The desktop agentos. arXiv preprint arXiv:2504.14603

    work page arXiv 2025

  40. [40]

    Zhang, J.; Yu, Y.; Liao, M.; Li, W.; Wu, J.; and Wei, Z. 2024. UI-Hawk: Unleashing the screen stream understanding for gui agents

    work page 2024

  41. [41]

    A Survey of Large Language Models

    Zhao, W. X.; Zhou, K.; Li, J.; Tang, T.; Wang, X.; Hou, Y.; Min, Y.; Zhang, B.; Zhang, J.; Dong, Z.; et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  42. [42]

    Zhao, Y.; Harrison, B.; and Yu, T. 2024. Dinodroid: Testing android apps using deep q-networks. ACM Transactions on Software Engineering and Methodology, 33(5): 1--24

    work page 2024

  43. [43]

    SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills

    Zheng, B.; Fatemi, M. Y.; Jin, X.; Wang, Z. Z.; Gandhi, A.; Song, Y.; Gu, Y.; Srinivasa, J.; Liu, G.; Neubig, G.; and Su, Y. 2025. SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills. arXiv:2504.07079

    work page internal anchor Pith review Pith/arXiv arXiv 2025