Pith. sign in

REVIEW 3 major objections 4 minor 3 cited by

MapAgent: Trajectory-Constructed Memory-Augmented Planning for Mobile Task Automation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Retrieving page memories distilled from past trajectories improves mobile task automation, with cross-app success of 0.350 in both English and Chinese.

desk verdict MapAgent packages known memory-augmented planning ideas into a clean system, but the evaluation may be undermined by test-set contamination in the memory database. read the letter →

arxiv 2507.21953 v1 pith:HFPZTMRB submitted 2025-07-29 cs.HC cs.AI

Yi Kong , Dianxi Shi , Guoli Yang , Zhang ke-di , Chenlin Huang , Xiaopeng Li , Songchang Jin This is my paper
classification cs.HCcs.AI
keywords mobiletaskautomationLLMagentsmemory-augmentedplanningGUItrajectorymemoryvectordatabaseretrievalcoarse-to-finedual-LLMexecutor
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that LLM-based mobile agents fail at real-world tasks because the models lack concrete knowledge of real app pages, and that injecting retrieved page-level memories from past trajectories into the planner closes that gap. MapAgent converts each page visited in an execution trajectory into a compact page chunk—a description, key UI elements, the action path that reached it, and a label—and stores chunks in a per-app vector database. At task time, a coarse-to-fine planner splits the task into app-specific subtasks, retrieves the top three matching pages per app, and uses them as in-context exemplars for fine-grained planning; a dual-LLM executor (Decision-maker plus Judge) then executes and monitors progress. Reported success on SPA-Bench cross-app tasks is 0.350 in both English and Chinese against best baselines of 0.200 and 0.100, and on CHOP it is 0.800 (English) and 0.700 (Chinese) overall. If the results hold with genuinely disjoint memory trajectories, memory augmentation is a practical way to give GUI agents app knowledge without retraining.

What carries the argument

The central object is the page chunk, a structured snapshot of one screen: page description, key UI elements, an action path describing how the agent arrived, and a short page label. Each page in a trajectory is summarized into a chunk by a multimodal LLM, embedded, and stored in a per-app collection in a vector database. Retrieval is the second mechanism: a subtask query is embedded and the top three chunks by cosine similarity are injected into the planner, letting the stored action path function as an in-context exemplar. The third mechanism is the dual-LLM executor, where a Decision-maker generates ReAct-style thoughts and actions and a Judge compares before-and-after observations to evaluate progress and recommend the next move, with a short-term memory unit carrying information across subtasks.

What would settle it

Build the page-memory database only from trajectories of tasks not present in SPA-Bench or CHOP (for example, a disjoint app set or a held-out split), rerun the same benchmarks, and compare the cross-app success rates; if 0.350 drops toward the 0.200 baseline, the reported gain is test-set lookup rather than generalizable memory augmentation.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that trajectory-derived page memory is a reusable knowledge source for mobile agents. A multimodal LLM summarizes each screen in a trajectory into a structured page chunk, capturing layout, functional elements, the path that led there, and a label; chunks are embedded and stored per app so retrieval never mixes pages across confusingly similar apps. Planning is coarse-to-fine: the task is decomposed into subtasks, scheduled to specific apps, and then each app's subtasks query the memory database for the top-k similar pages, whose action paths serve as few-shot exemplars in the fine-grained prompt. Execution is handled by two collaborating LLMs, one that acts and one that judges progress before each next step. The result is a reported 0.350 cross-app success rate on SPA-Bench in both languages—double the 0.200 English and more than triple the 0.100 Chinese best baselines—and 0.800/0.700 overall on CHOP, the highest in the comparison.

Load-bearing premise

The reported gains depend on the trajectories used to build the page memory being distinct from the benchmark test tasks, because otherwise the retrieved pages hand the agent the exact action path rather than testing generalization.

Editorial extensions

If this is right

  • Cross-app tasks benefit most, because coarse-to-fine decomposition keeps context windows small while retrieved pages and short-term memory connect the subtasks.
  • Retrieved action paths act as few-shot exemplars, which the ablation shows improves success at every difficulty level rather than only on easy tasks.
  • The Judge component adds an extra LLM call per step but buys its cost on multi-step tasks: success jumps at SPA-Bench difficulty Levels 2 and 3 when it is enabled.
  • The per-app memory collections prevent cross-app confusion between visually similar elements, so memory reuse is organized by application rather than by raw similarity alone.
  • The reported token and time costs stay near the lower end of the comparison, suggesting the memory overhead is retrieval and one extra planning round rather than repeated exploration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the memory trajectories are indeed disjoint from benchmark tasks, the same machinery should transfer to new apps after a single exploration pass, but the paper does not specify how many trajectories per app are needed or what the cold-start cost is.
  • A natural extension the paper leaves implicit is updating the memory database from successful executions at runtime, turning the static trajectory collection into a self-improving agent loop.
  • The hard-task gaps on CHOP (0.6 English, 0.5 Chinese) suggest the ceiling is not memory alone; testing whether retrieving failure trajectories or adding error corrections lifts hard-task success would isolate the next bottleneck.
  • Because memories are stored as text chunks, one could probe cross-lingual transfer—whether English pages retrieved for an English subtask help a Chinese version of the same task—which the paper does not test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes MapAgent, an LLM-based mobile GUI automation agent that augments task planning with a trajectory-constructed page-memory database. A memory mechanism summarizes pages from historical trajectories into structured page chunks (page description, key UI elements, action path, page label), stores them per app in a vector database, and retrieves top-k pages by cosine similarity during a coarse-to-fine planning phase. A dual-role executor (Decision-maker and Judge) then executes the plan. Experiments on SPA-Bench and CHOP report success-rate improvements over several baselines in English and Chinese, with ablations and cost analyses. The authors conclude that memory-augmented planning materially improves performance.

Significance. If the reported results hold, MapAgent would be a useful contribution to mobile task automation: the memory design is simple, the idea of reusing page-level action paths as in-context exemplars is plausible, and the real-device evaluation across two benchmarks and two languages is a strength. The paper also provides ablations, an error analysis, and cost measurements. However, the central empirical claim currently rests on an unresolved separation between the trajectories used to build memory and the benchmark tasks used for evaluation, and the quantitative comparisons lack variance information. These issues must be addressed before the generalization claim can be accepted. If the memory database is later shown to be built from held-out trajectories and the statistical concerns are handled, the contribution would be solidly incremental and useful for the GUI-agent community.

major comments (3)
  1. [Trajectory-based Memory Mechanism; Eqs. (5)-(6)] The paper does not state whether the trajectories used to construct the page-memory database are disjoint from the SPA-Bench and CHOP evaluation tasks. Each stored page chunk contains an 'Action Path' describing the route from the initial page to the current page, and the fine-grained planner is instructed with retrieved pages whose action paths 'serve as an exemplar' (Ablation Study). If the exploration tasks that generated the trajectories overlap with the benchmark tasks, then the retrieved context for a test task can be nearly the exact solution, and the reported gains (e.g., 0.350 vs. 0.200/0.100 on SPA-Bench cross-app in Table 1) would reflect test-set lookup rather than planning generalization. The ablation in Table 3 cannot resolve this concern because every 'with memory' configuration may use the same potentially overlapping memory. To support the central claim, the authors must either state and justify a split between memory-building tasks and evaluation tasks, or rerun the experiments with memory restricted to trajectories generated from held-out tasks.
  2. [Experiment Setup, Benchmark] The benchmark description is internally inconsistent: 'SPA-Bench consists of 340 tasks, including 150 single-app tasks and 20 cross-app tasks' does not add up, and no third task category is described. This matters for interpreting Table 1, because if the cross-app setting contains only 20 tasks total, a 0.350 vs. 0.200 difference corresponds to a very small number of tasks. Please give the exact total number of tasks, the per-language and per-category denominators for both SPA-Bench and CHOP, and clarify whether the listed numbers are a subset of the full benchmark.
  3. [Experimental Results, Tables 1-2] No variance, confidence intervals, or significance tests are reported. The per-cell sample sizes are small; for example, in CHOP each difficulty-level cell appears to contain 10 tasks, so a 0.1 difference in success rate is one task. This is especially relevant where the claimed advantage is not visually supported, such as SPA-Bench single-app English (0.627 vs. 0.640 for M3A) and CHOP English hard tasks (0.6 for both CHOP and MapAgent). The authors should report repeated runs or a statistical test so that the reader can assess whether the reported differences are within noise.
minor comments (4)
  1. [Ablation Study, Table 3] The sentence 'when only memory is disabled (w/o M), there is a corresponding increase in success rate' is confusing: relative to 'w/o M & J' the success rate does increase, but relative to the full MapAgent it decreases. Please specify the reference configuration explicitly.
  2. [Figure 3] The figure caption and the axis labels are unclear about which quantity is plotted on each axis, and the units appear inconsistent ('Mean Token Cost per step (usd)' appears to be used for both the vertical axis and the plotted-point labels, while the horizontal axis is labeled in seconds). Please clarify the axes, units, and whether the plotted values are means over all single-app tasks.
  3. [Related Work and References] The reference for M3A is listed as 'AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents', which does not appear to be the M3A paper used as a baseline; please correct this citation so that the baseline is properly identified.
  4. [Method, Task Executor] The term 'dual-LLM architecture' is used even though both roles are implemented by the same GPT-4o model; please clarify whether two distinct model instances or two prompt-based roles within one model are intended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: retrieval of historical page memory is the method's intended mechanism, not a hidden fit; missing train/test split is an evaluation risk, not circularity.

full rationale

MapAgent is an empirical retrieval-augmented agent paper; it contains no analytic derivation that could reduce to its inputs. The trajectory-to-page-memory construction described in 'Trajectory-based Memory Mechanism' and the retrieval equations (Eqs. 5-6) implement the intended method: past pages and action paths are injected into the LLM planner as in-context exemplars. This is the mechanism under evaluation, not a circular prediction. The SPA-Bench and CHOP results are external benchmarks, and the ablation in Table 3 separately isolates the memory component and the Judge component, providing independent evidence that the memory mechanism contributes to the reported success rates. The one significant caveat is that the paper never explicitly states whether the historical trajectories used to build the page-memory database are disjoint from the benchmark tasks; if they overlapped, some reported gains could reflect retrieving near-identical action sequences rather than generalizable planning. However, this is a data-hygiene and evaluation-validity risk, not a demonstrated circularity: there is no equation, fitted parameter, or cited prior result that encodes the benchmark outcomes into the memory by construction. The paper also shows a concrete case (Fig. 4) where memory resolves a task that failed without memory, illustrating the intended transfer mechanism. There are no load-bearing self-citations, no imported uniqueness theorems, no fitted values renamed as predictions, and no known result merely renamed. Accordingly, the central claim retains independent empirical content and the circularity score is 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the fidelity of page summaries, the correctness of retrieval, the assumption that memory construction is fair, and the assumption that decomposing tasks per app does not lose cross-app intent. The only tunable number is the retrieval count.

free parameters (1)
  • top-k retrieval count = 3
    The fine-grained planner retrieves the top 3 most similar pages per subtask (Eq. 5). The value is chosen without a sensitivity analysis.
assumptions (4)
  • domain assumption GPT-4o as a multimodal summarizer produces accurate page chunks (description, key UI elements, action path, label)
    Used in Trajectory-based Memory Mechanism; errors in summarization propagate directly to retrieval and planning.
  • domain assumption Historical trajectories used to build memory are disjoint from benchmark test tasks
    Needed for the reported gains to reflect generalization rather than test-set memorization; the paper never states this.
  • domain assumption Cosine similarity between embedded subtask queries and page vectors retrieves task-relevant pages
    Eq. (5) in Memory-Augmented Task Planning; no validation of retrieval quality is provided.
  • domain assumption Coarse-to-fine decomposition into per-app subtasks preserves the semantics of the original user task
    The Task Scheduler (Eq. 4) assigns subtasks to apps, and the executor only sees per-app plans; a wrong split loses the cross-app goal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MapAgent: Trajectory-Constructed Memory-Augmented Planning for Mobile Task Automation." pith.science (2026). https://pith.science/paper/HFPZTMRB

@misc{pith2026250721953,
  author       = {Pith},
  title        = {Pith review of: MapAgent: Trajectory-Constructed Memory-Augmented Planning for Mobile Task Automation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HFPZTMRB}},
  note         = {Machine review of arXiv:2507.21953}
}
read the original abstract

The recent advancement of autonomous agents powered by Large Language Models (LLMs) has demonstrated significant potential for automating tasks on mobile devices through graphical user interfaces (GUIs). Despite initial progress, these agents still face challenges when handling complex real-world tasks. These challenges arise from a lack of knowledge about real-life mobile applications in LLM-based agents, which may lead to ineffective task planning and even cause hallucinations. To address these challenges, we propose a novel LLM-based agent framework called MapAgent that leverages memory constructed from historical trajectories to augment current task planning. Specifically, we first propose a trajectory-based memory mechanism that transforms task execution trajectories into a reusable and structured page-memory database. Each page within a trajectory is extracted as a compact yet comprehensive snapshot, capturing both its UI layout and functional context. Secondly, we introduce a coarse-to-fine task planning approach that retrieves relevant pages from the memory database based on similarity and injects them into the LLM planner to compensate for potential deficiencies in understanding real-world app scenarios, thereby achieving more informed and context-aware task planning. Finally, planned tasks are transformed into executable actions through a task executor supported by a dual-LLM architecture, ensuring effective tracking of task progress. Experimental results in real-world scenarios demonstrate that MapAgent achieves superior performance to existing methods. The code will be open-sourced to support further research.

Figures

Figures reproduced from arXiv: 2507.21953 by the authors.

Figure 1
Figure 1. Overview of MapAgent framework. (a) We employ a multimodal LLM to summarize each page of a trajectory within [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. When the task Check the details of General settings in Google Play is completed, taking the last page in this execution trajectory as an example, all tasks related to the UI Elements of this page can follow the same execution path. and successfully reached the target page, it still needs to un￾dergo the same trial-and-error process when navigating this page again. This behavior is analogous to a person using a mobil… view at source ↗
Figure 3
Figure 3. Computational overhead analysis. Mean Time Spent per step (MTS) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: An example of a cross-app task planning process. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. MobiBench: Multi-Branch, Modular Benchmark for Mobile GUI Agents

    cs.AI 2025-12 conditional novelty 8.0 of 10

    MobiBench reaches near-human offline evaluation fidelity for mobile GUI agents by accepting any valid action at each step, and enables modular attribution of performance to agent components.

  2. Externalizing Research Synthesis and Validation in AI Scientists through a Research Harness

    cs.AI 2026-06 unverdicted novelty 6.0 of 10

    Xcientist is a research harness that externalizes an AI scientist's literature grounding, idea evolution, experiments, and repairs into auditable artifacts, demonstrated on memory, traffic forecasting, and PDE-solving tasks.

  3. ActMem: Bridging the Gap Between Memory Retrieval and Reasoning in LLM Agents

    cs.CL 2026-02 conditional novelty 6.0 of 10

    ActMem stores agent dialogue as a causal and semantic knowledge graph and uses LLM counterfactual reasoning to retrieve implicit constraints, outperforming RAG-style baselines on a new reasoning benchmark.

Reference graph

Works this paper leans on

59 extracted references · 33 canonical work pages · cited by 3 Pith papers

  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...

  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...

  3. [3]

    B.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J.; Dhariwal, P.; et al

    Brown, T. B.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J.; Dhariwal, P.; et al. 2020. Language models are few-shot learners. In Proceedings of the 34th International Conference on Neural Information Processing Systems

    work page 2020

  4. [4]

    Burns, A.; Arsan, D.; Agrawal, S.; Kumar, R.; Saenko, K.; and Plummer, B. A. 2022. A Dataset for Interactive Vision Language Navigation with Unknown Command Feasibility. In European Conference on Computer Vision

    work page 2022

  5. [5]

    Chen, J.; Lin, B.; Xu, R.; Chai, Z.; Liang, X.; and Wong, K.-Y. 2024. M ap GPT : Map-Guided Prompting with Adaptive Path Planning for Vision-and-Language Navigation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics

    work page 2024

  6. [6]

    Chen, J.; Yuen, D.; Xie, B.; Yang, Y.; Chen, G.; Wu, Z.; et al. 2025. SPA-Bench: A Comprehensive Benchmark for SmartPhone Agent Evaluation. In The Thirteenth International Conference on Learning Representations

    work page 2025

  7. [7]

    DeepSeek-AI; Liu, A.; Feng, B.; Xue, B.; Wang, B.; Wu, B.; et al. 2025. DeepSeek-V3 Technical Report. arXiv:2412.19437

    arXiv 2025

  8. [8]

    Gou, B.; Wang, R.; Zheng, B.; Xie, Y.; Chang, C.; Shu, Y.; Sun, H.; and Su, Y. 2025. Navigating the Digital World as Humans Do: Universal Visual Grounding for GUI Agents. In The Thirteenth International Conference on Learning Representations

    2025

Show all 59 references
  1. [9]

    U.; Qureshi, R.; Shah, A.; Irfan, M.; Zafar, A.; Shaikh, M

    Hadi, M. U.; Qureshi, R.; Shah, A.; Irfan, M.; Zafar, A.; Shaikh, M. B.; et al. 2023. A survey on large language models: Applications, challenges, limitations, and practical usage. Authorea Preprints

    2023

  2. [10]

    Hong, S.; Zhuge, M.; Chen, J.; Zheng, X.; Cheng, Y.; Wang, J.; et al. 2024. Meta GPT : Meta Programming for A Multi-Agent Collaborative Framework. In The Twelfth International Conference on Learning Representations

    2024

  3. [11]

    Hu, X.; Xiong, T.; Yi, B.; Wei, Z.; Xiao, R.; Chen, Y.; et al. 2025. OS Agents: A Survey on MLLM -based Agents for Computer, Phone and Browser Use. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics

    2025

  4. [12]

    Huang, L.; Yu, W.; Ma, W.; Zhong, W.; Feng, Z.; Wang, H.; et al. 2025. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ACM Transactions on Information Systems

    2025

  5. [13]

    P.; Perelman, A.; Ramesh, A.; Clark, A.; et al

    Hurst, A.; Lerer, A.; Goucher, A. P.; Perelman, A.; Ramesh, A.; Clark, A.; et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  6. [14]

    S.; Xu, X.; McDuff, D.; et al

    Kim, Y.; Park, C.; Jeong, H.; Chan, Y. S.; Xu, X.; McDuff, D.; et al. 2024. Mdagents: An adaptive collaboration of llms for medical decision-making. In Advances in Neural Information Processing Systems

    2024

  7. [15]

    Li, D.; Jiang, B.; Huang, L.; Beigi, A.; Zhao, C.; Tan, Z.; Bhattacharjee, A.; Jiang, Y.; Chen, C.; Wu, T.; Shu, K.; Cheng, L.; and Liu, H. 2025 a . From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge. arXiv:2411.16594

    2025

  8. [16]

    Li, G.; Hammoud, H. A. A. K.; Itani, H.; Khizbullin, D.; and Ghanem, B. 2023. CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society. In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  9. [17]

    Li, J.; Lai, Y.; Li, W.; Ren, J.; Zhang, M.; Kang, X.; Wang, S.; Li, P.; Zhang, Y.-Q.; Ma, W.; and Liu, Y. 2025 b . Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents. arXiv:2405.02957

    2025 arXiv

  10. [18]

    Li, Y.; He, J.; Zhou, X.; Zhang, Y.; and Baldridge, J. 2020. Mapping Natural Language Instructions to Mobile UI Action Sequences. In Annual Conference of the Association for Computational Linguistics

    2020

  11. [19]

    Li, Y.; and Riva, O. 2021. Glider: A reinforcement learning approach to extract UI scripts from websites. In 44th International ACM SIGIR Conference on Research and Development in Information Retrieval

    2021

  12. [20]

    Liu, B.; Li, X.; Zhang, J.; Wang, J.; He, T.; Hong, S.; et al. 2025 a . Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems. arXiv:2504.01990

    2025 arXiv

  13. [21]

    Liu, G.; Zhao, P.; Liu, L.; Guo, Y.; Xiao, H.; Lin, W.; et al. 2025 b . LLM-Powered GUI Agents in Phone Automation: Surveying Progress and Prospects. arXiv:2504.19838

    2025

  14. [22]

    Min, S.; Lyu, X.; Holtzman, A.; Artetxe, M.; Lewis, M.; Hajishirzi, H.; and Zettlemoyer, L. 2022. Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing

    2022

  15. [23]

    OpenAI. 2023. GPT-4V(ision) System Card

    2023

  16. [24]

    OpenAI; Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; et al. 2024. GPT-4 Technical Report. arXiv:2303.08774

    2024 arXiv

  17. [25]

    S.; O'Brien, J.; Cai, C

    Park, J. S.; O'Brien, J.; Cai, C. J.; Morris, M. R.; Liang, P.; and Bernstein, M. S. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology

    2023

  18. [26]

    G.; Zhang, T.; Wang, X.; and Gonzalez, J

    Patil, S. G.; Zhang, T.; Wang, X.; and Gonzalez, J. E. 2024. Gorilla: Large Language Model Connected with Massive APIs. In Advances in Neural Information Processing Systems

    2024

  19. [27]

    Qin, Y.; Ye, Y.; Fang, J.; Wang, H.; Liang, S.; Tian, S.; et al. 2025. UI-TARS: Pioneering Automated GUI Interaction with Native Agents. arXiv:2501.12326

    2025 arXiv

  20. [28]

    Qin, Y.; Zhou, E.; Liu, Q.; Yin, Z.; Sheng, L.; Zhang, R.; Qiao, Y.; and Shao, J. 2024. MP5: A Multi-modal Open-ended Embodied System in Minecraft via Active Perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  21. [29]

    Rawles, C.; Clinckemaillie, S.; Chang, Y.; Waltz, J.; Lau, G.; Fair, M.; et al. 2025. AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents. In The Thirteenth International Conference on Learning Representations

    2025

  22. [30]

    Shinn, N.; Cassano, F.; Gopinath, A.; Narasimhan, K.; and Yao, S. 2023. Reflexion: language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems

    2023

  23. [31]

    Significant-Gravitas. 2023. AutoGPT

    2023

  24. [32]

    Tan, W.; Zhang, W.; Xu, X.; Xia, H.; Ding, Z.; Li, B.; et al. 2024. Cradle: Empowering Foundation Agents Towards General Computer Control. arXiv:2403.03186

    2024 arXiv

  25. [33]

    Tang, J.; Gao, H.; Pan, X.; Wang, L.; Tan, H.; Gao, D.; et al. 2024. GenSim: A General Social Simulation Platform with Large Language Model based Agents. arXiv:2410.04360

    2024 arXiv

  26. [34]

    Tao, M.; Zhao, D.; and Feng, Y. 2025. Chain-of-Discussion: A Multi-Model Framework for Complex Evidence-Based Question Answering. In Proceedings of the 31st International Conference on Computational Linguistics

    2025

  27. [35]

    Toyama, D.; Hamel, P.; Gergely, A.; Comanici, G.; Glaese, A.; Ahmed, Z.; et al. 2021. AndroidEnv: A Reinforcement Learning Platform for Android. arXiv:2105.13231

    2021 arXiv

  28. [36]

    Wang, J.; Xu, H.; Jia, H.; Zhang, X.; Yan, M.; Shen, W.; et al. 2024 a . Mobile-Agent-v2: Mobile Device Operation Assistant with Effective Navigation via Multi-Agent Collaboration. In Advances in Neural Information Processing Systems

    2024

  29. [37]

    Wang, J.; Yi, X.; Guo, R.; Jin, H.; Xu, P.; Li, S.; et al. 2021. Milvus: A Purpose-Built Vector Data Management System. In Proceedings of the 2021 International Conference on Management of Data

    2021

  30. [38]

    Wang, L.; Ma, C.; Feng, X.; Zhang, Z.; Yang, H.; Zhang, J.; et al. 2024 b . A survey on large language model based autonomous agents. Frontiers of Computer Science

    2024

  31. [39]

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Ichter, B.; Xia, F.; et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems

    2022

  32. [40]

    J.-J.; et al

    Wen, H.; Li, Y.; Liu, G.; Zhao, S.; Yu, T.; Li, T. J.-J.; et al. 2024. AutoDroid: LLM-powered Task Automation in Android. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking

    2024

  33. [41]

    Wu, Q.; Liu, W.; Luan, J.; and Wang, B. 2025. ReachAgent: Enhancing Mobile Agent via Page Reaching and Operation. arXiv:2502.02955

    2025 arXiv

  34. [42]

    Wu, Z.; Han, C.; Ding, Z.; Weng, Z.; Liu, Z.; Yao, S.; et al. 2024. OS -Copilot: Towards Generalist Computer Agents with Self-Improvement. In ICLR 2024 Workshop on Large Language Model (LLM) Agents

    2024

  35. [43]

    Yang, J.; Zhang, H.; Li, F.; Zou, X.; Li, C.; and Gao, J. 2023. Set-of-Mark Prompting Unleashes Extraordinary Visual Grounding in GPT-4V. arXiv:2310.11441

    2023 arXiv

  36. [44]

    R.; and Cao, Y

    Yao, S.; Zhao, J.; Yu, D.; Du, N.; Shafran, I.; Narasimhan, K. R.; and Cao, Y. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In The Eleventh International Conference on Learning Representations

    2023

  37. [45]

    Zawalski, M.; Chen, W.; Pertsch, K.; Mees, O.; Finn, C.; and Levine, S. 2025. Robotic Control via Embodied Chain-of-Thought Reasoning. arXiv:2407.08693

    2025 arXiv

  38. [46]

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

    2025 arXiv

  39. [47]

    Zhang, C.; He, S.; Qian, J.; Li, B.; Li, L.; Qin, S.; et al. 2025 b . Large Language Model-Brained GUI Agents: A Survey. arXiv:2411.18279

    2025 arXiv

  40. [48]

    Zhang, C.; Yang, Z.; Liu, J.; Li, Y.; Han, Y.; Chen, X.; et al. 2025 c . AppAgent: Multimodal Agents as Smartphone Users. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems

    2025

  41. [49]

    Zhang, D.; Yu, Y.; Dong, J.; Li, C.; Su, D.; Chu, C.; and Yu, D. 2024 a . Mm-llms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601

    2024 arXiv

  42. [50]

    Zhang, J.; Zhao, C.; Zhao, Y.; Yu, Z.; He, M.; and Fan, J. 2024 b . MobileExperts: A Dynamic Tool-Enabled Agent Team in Mobile Devices. arXiv:2407.03913

    2024 arXiv

  43. [51]

    Zhang, X.; Guo, J.; Zhao, S.; Fu, M.; Duan, L.; Hu, J.; et al. 2025 d . Unified Multimodal Understanding and Generation Models: Advances, Challenges, and Opportunities. arXiv:2505.02567

    2025

  44. [52]

    Zhang, Y.; Li, Y.; Cui, L.; Cai, D.; Liu, L.; Fu, T.; et al. 2023. Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models. arXiv:2309.01219

    2023 arXiv

  45. [53]

    X.; Zhou, K.; Li, J.; Tang, T.; Wang, X.; et al

    Zhao, W. X.; Zhou, K.; Li, J.; Tang, T.; Wang, X.; et al. 2025. A Survey of Large Language Models. arXiv:2303.18223

    2025 arXiv

  46. [54]

    Zheng, B.; Gou, B.; Kil, J.; Sun, H.; and Su, Y. 2024. GPT-4V(ision) is a Generalist Web Agent, if Grounded. In Forty-first International Conference on Machine Learning

    2024

  47. [55]

    Zheng, L.; Chiang, W.-L.; Sheng, Y.; Zhuang, S.; Wu, Z.; et al. 2023. Judging LLM-as-a-judge with MT-bench and Chatbot Arena. In Proceedings of the 37th International Conference on Neural Information Processing Systems

    2023

  48. [56]

    Zhou, Y.; Wang, S.; Dai, S.; Jia, Q.; Du, Z.; Dong, Z.; and Xu, J. 2025. CHOP: Mobile Operating Assistant with Constrained High-frequency Optimized Subtask Planning. arXiv:2503.03743

    2025 arXiv

  49. [57]

    Zhu, Z.; Tang, H.; Li, Y.; Liu, D.; Xu, H.; Lan, K.; et al. 2025. MOBA: Multifaceted Memory-Enhanced Adaptive Planning for Efficient Mobile Task Automation. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Lingui...

    2025

  50. [58]

    A.; et al

    Zhuang, Y.; Chen, X.; Yu, T.; Mitra, S.; Bursztyn, V.; Rossi, R. A.; et al. 2024. ToolChain*: Efficient Action Space Navigation in Large Language Models with A* Search. In The Twelfth International Conference on Learning Representations

    2024

  51. [59]

    Zhuge, M.; Wang, W.; Kirsch, L.; Faccio, F.; Khizbullin, D.; and Schmidhuber, J. 2024. GPTSwarm: Language Agents as Optimizable Graphs. In Forty-first International Conference on Machine Learning

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.