Pith. sign in

REVIEW 3 major objections 5 minor 4 cited by

AppVLM: A Lightweight Vision Language Model for Online App Control

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A 3-billion-parameter vision-language model matches GPT-4o agents at smartphone app control while running up to ten times faster.

desk verdict Solid ReST-style recipe on a 3B VLM, but the abstract's 'highest accuracy' and 'first lightweight' claims don't survive contact with its own tables. read the letter →

arxiv 2502.06395 v1 pith:IRKN7M54 submitted 2025-02-10 cs.AI

Georgios Papoudakis , Thomas Coste , Zhihao Wu , Jianye Hao , Jun Wang , Kun Shao This is my paper · ORCID
classification cs.AI
keywords appagentsvision-languagemodelAndroidWorldControlreinforcedself-trainingsmartphoneautomationGUIefficientinference
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

This paper claims that a lightweight 3B vision-language model can be trained to control Android apps well enough to match large proprietary GPT-4o agents on the AndroidWorld benchmark, while running up to ten times faster. The training recipe starts with supervised fine-tuning on the AndroidControl dataset, then alternates between collecting successful trajectories in the AndroidWorld emulator and fine-tuning on them—a loop the paper calls Reinforce Fine-Tuning (RFT). In offline action-prediction tests on AndroidControl, the SFT-only model (AppVLM-base) reaches 73.9% accuracy on the in-distribution split, the highest among the evaluated baselines. Online, the final AppVLM solves 37.8% of 82 AndroidWorld tasks on average, compared with 41.9% for the best GPT-4o baseline (T3A) and 36.6% for M3A, with an average inference time of 0.91 seconds per action. The paper argues this makes a practical, low-cost alternative to querying large proprietary models.

What carries the argument

The load-bearing mechanism is the Reinforce Fine-Tuning (RFT) loop, a version of reinforced self-training: the current policy is deployed in AndroidWorld emulators through a distributed client-server queue, successful trajectories are collected at high sampling temperatures, redundant steps are removed by comparing consecutive screenshots, under-solved tasks are oversampled up to a threshold of ten successes, and the policy is fine-tuned to maximize the log-likelihood of the successful actions. The observation format matters too: each screenshot is annotated with numbered bounding boxes of clickable UI elements from the accessibility tree, and the textual input contains the goal plus the last five actions, with click targets stored as condensed UI-element descriptions. A final full SFT of the original AndroidControl-trained base model on all collected data combines offline accuracy with online competence.

What would settle it

Re-run T3A and M3A on the same 82-task AndroidWorld subset with the same three-seed protocol used for AppVLM. If their success rates differ from the reported 41.9% and 36.6%, the claimed parity with GPT-4o is not established; a complementary check is reproducing AppVLM-base's 73.9% AndroidControl IDD accuracy under the paper's relaxed matching rule.

Watch

Extended reading notes

Core claim

The central claim is that a lightweight VLM, AppVLM, can close most of the performance gap with GPT-4o agents for online Android app control, and actually surpass them on offline action prediction, because its two-stage training transfers from a static human-demonstration dataset to a dynamic emulator environment. AppVLM-base—Paligemma-3B-896 fine-tuned only on AndroidControl—attains 73.9% relaxed action accuracy on the IDD AndroidControl split and 65.4% on the app-unseen split, beating the previous best fine-tuned baseline (LT-all-r64) and all GPT-4o baselines. Adding RFT (three iterations of collecting successful AndroidWorld trajectories, filtering redundant steps, oversampling under-solved tasks, and fine-tuning with maximum likelihood on successful trajectories) raises online success from 11.4% to 37.8%, then a final SFT step on the original base model recovers offline accuracy while keeping online gains. AppVLM solves 37.8% of tasks on the paper's 82-task subset, exceeding M3A (36.6%) and SeeAct (22.0%), and coming 4 percentage points behind T3A (41.9%), while taking 0.91 seconds per action versus 4.29–15.82 seconds for GPT-4o baselines.

Load-bearing premise

The comparison assumes the GPT-4o baseline success rates (T3A 41.9%, M3A 36.6%) were measured on the same 82-task AndroidWorld subset and under the same evaluation protocol as AppVLM, something the paper does not explicitly confirm.

Editorial extensions

If this is right

  • An SFT-only 3B model can already set a higher offline action-prediction accuracy on AndroidControl than GPT-4o prompt-based agents and prior fine-tuned baselines.
  • The RFT loop produces a large online gain (11.4% to 37.8% AndroidWorld success) without sacrificing offline performance when followed by a final SFT on the base model.
  • AppVLM's per-action inference time (0.91 seconds) is roughly five times faster than T3A and over ten times faster than M3A and SeeAct, making the success rates achievable at a fraction of GPT-4o's cost.
  • Higher AndroidControl action accuracy does not automatically imply higher AndroidWorld success: Llama-3 and T3A outperform their visual counterparts offline yet underperform online.
  • The model can acquire rare or absent behaviors during RFT, such as long-press (under 1% of AndroidControl actions), once such actions appear in successful collected trajectories.

Reading between the lines

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

  • The headline comparison to GPT-4o would be fully like-for-like only if the T3A/M3A numbers were rerun on the same 82-task subset with the same seed protocol; the paper does not explicitly say so, and the reported 4-point gap could shift if those baselines were drawn from the full 116-task benchmark.
  • If the efficiency holds at deployment scale, app agents could run locally on users' devices, shifting the practical bottleneck from inference cost to task coverage—the paper itself notes failures on unseen operations like clipboard use.
  • A dedicated, robust reward model for app control would let the RFT loop scale beyond AndroidWorld's 116 tasks, since the current environment is the only one with an internal reward signal.
  • A direct testable extension is to start RFT from a different base VLM (the paper suggests Llama-3 as an alternative) and check whether the same AppVLM-base-plus-RFT improvement curve appears.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes AppVLM, a 3B-parameter vision-language model for Android app control. The method first performs supervised fine-tuning (SFT) on the AndroidControl dataset (producing AppVLM-base), then iteratively collects successful trajectories from the AndroidWorld environment and fine-tunes the policy (Reinforce Fine-Tuning, RFT), and finally performs an additional SFT step on AppVLM-base using all collected AndroidWorld data. The paper reports AndroidControl action-prediction accuracy across four splits and AndroidWorld online success rates across difficulty levels, and claims that AppVLM achieves state-of-the-art offline accuracy, matches GPT-4o baselines online, and runs up to ten times faster than GPT-4o-based agents.

Significance. If the empirical claims hold after the necessary corrections, the work would be a practical contribution: it demonstrates that a small open VLM can be made competitive with large proprietary GPT-4o agents on AndroidWorld at a fraction of the inference cost. The paper's strengths include the distributed client-server data-collection pipeline, the clear description of the RFT/ReST procedure, useful ablations in Tables 3 and 4, and a candid Section 4.5 that acknowledges that AppVLM-base, not the final AppVLM, achieves the best AndroidControl accuracy. However, the advertised claims in the abstract and contribution list outrun the data in Table 1, and the online comparison with GPT-4o baselines requires explicit clarification of evaluation protocols.

major comments (3)
  1. [Abstract; Section 1; Table 1] The statement that 'AppVLM achieves the highest action prediction accuracy in offline evaluation on the AndroidControl dataset' is contradicted by Table 1. AppVLM-base achieves 73.9/65.9/65.1/65.4 on IDD/Task-Unseen/Cat-Unseen/App-Unseen, while the final AppVLM achieves only 69.0/62.7/61.9/62.2, and the LT-all-r64 baseline reports 70.8 on IDD. Since Algorithm 1 defines AppVLM as the output of the final SFT step, the named model is not the state-of-the-art. The abstract and the first contribution bullet should be revised to attribute the AndroidControl SOTA to AppVLM-base, or the terminology in Algorithm 1 should be changed so that 'AppVLM' refers to the whole pipeline rather than the final checkpoints.
  2. [Section 4.4; Table 2; Appendix A.3] The paper does not state whether the GPT-4o baseline results (T3A, M3A, SeeAct) in Table 2 were obtained on the same 82-task subset used for AppVLM, under the same evaluation protocol and seeds. Appendix A.3 shows that the 82-task subset has a harder difficulty distribution than the full 116-task benchmark (46.3% vs 52.6% easy). If the baseline numbers were taken from Rawles et al. (2024) on the full benchmark, then the claim that AppVLM 'matches GPT-4o' is not a like-for-like comparison. The authors should either re-run the baselines on their subset or clearly state the provenance of every number in Table 2, including the number of seeds and any filtering applied.
  3. [Introduction; Section 4.5; Section 3.3] The claim that RFT provides 'adaptability to out-of-distribution tasks' and enables the agent to 'tackle tasks beyond those encountered during the initial SFT' is not established by the reported experiments. The RFT data are collected from AndroidWorld, and the online evaluation is also run on AndroidWorld with randomized parameters; the AndroidControl OOD splits are still drawn from the same dataset distribution as the SFT data. To support the OOD generalization claim, the authors should evaluate on tasks or apps that were not used during RFT data collection, or temper the claim to 'generalization to unseen parameterizations of AndroidWorld tasks.'
minor comments (5)
  1. [Section 4.1] There is a duplicated article in the sentence 'This procedure continues until either the the task is solved'.
  2. [References] The reference to 'Willams, R. J.' should be 'Williams, R. J.'.
  3. [Section 3.3.1] The threshold tau is described as 'solved fewer than a threshold tau times' and later defined as tau = 10; please clarify whether tau is the target number of successes used for oversampling, and state how the oversampling frequency is computed from tau.
  4. [Appendix A.1] The sentence 'AndroidControl also includes a click target as part of its input-text action, while we choose to keep these as separate actions as in AndroidControl' is confusing; please reword to clearly state the difference between the AndroidControl action format and the AppVLM action space.
  5. [Table 2] The average inference times for AppVLM and AppVLM-base are both 0.91 seconds; consider adding a note that these are the same model architecture and that the value is the average per action-generation step.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical benchmark with explicit SFT/RFT; abstract/Table 1 mismatch is a correctness issue, not a circular reduction.

full rationale

AppVLM is an empirical systems paper; its contributions are measured accuracies and success rates, not analytic derivations. The central pipeline (SFT on AndroidControl, then ReST/RFT on AndroidWorld trajectories, then final SFT) is described with explicit objectives (Eq. 1) and Algorithm 1; no prediction is produced by fitting a parameter to the same quantity it is said to predict. Evaluation on AndroidControl held-out splits is standard supervised evaluation; online AndroidWorld evaluation, while conducted in the same environment used for RFT data collection, is standard RL/ReST evaluation and the paper does not present it as out-of-distribution with respect to RFT. The self-citations (LiMAC, DistRL) appear only in related work and are not load-bearing. The abstract's claim that 'AppVLM' achieves the highest AndroidControl accuracy is internally inconsistent with Table 1, where AppVLM-base is strictly higher on all four splits, but that is a reporting/correctness inconsistency rather than a circular step: no claim reduces to its own input by definition or by a fitted parameter. Likewise, whether the GPT-4o baselines were re-run on the same 82-task AndroidWorld subset is a measurement-comparability concern, not circularity. No patterns from the enumerated list are present, so the honest finding is no significant circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central performance claims rest on the AndroidWorld evaluation protocol and the quality of the training data. The paper introduces no new mathematical entities, but it does depend on several domain assumptions about the environment, the dataset, and the effectiveness of self-training. The hyperparameters listed below are chosen by hand and not swept, so they are free parameters in the practical sense.

free parameters (7)
  • learning rate = 3e-6
    Used for all SFT and RFT fine-tuning; no sweep reported.
  • number of epochs = 3
    Fixed for all fine-tuning steps.
  • effective batch size = 64
    Fixed for all fine-tuning.
  • oversampling threshold tau = 10
    Tasks solved fewer than tau times are oversampled to match frequent tasks; chosen by hand.
  • action history length = 5
    The five most recent actions are included as text; chosen to limit tokens.
  • number of RFT iterations = 3 (text) / 4 (Table 3)
    The text states three iterations, but Table 3 reports four RFT checkpoints; inconsistency in the paper.
  • data collection sampling temperature = high (unspecified)
    Used to encourage diverse trajectories; exact value not reported.
assumptions (4)
  • domain assumption AndroidWorld provides a valid reward signal and the 82-task subset is a fair basis for comparing agents.
    The paper removes 34 of 116 tasks (verification, Q&A, drawing) and argues the subset is harder, but this changes the evaluation distribution relative to published baselines.
  • domain assumption Successful AndroidWorld trajectories collected by the policy are a suitable training signal for improving generalization.
    RFT relies on rejection sampling of successful rollouts; if success is sparse or the trajectory has redundant steps, the filter only removes unchanged screenshots and may reinforce suboptimal actions.
  • domain assumption AndroidControl human demonstrations are near-optimal and the relaxed action-accuracy metric is a valid proxy for action correctness.
    The paper uses relaxed click target matching (bounding box within target element) and converts coordinates to indices, which may introduce annotation noise.
  • domain assumption Paligemma-3B-896 is a suitable base model for app control.
    The choice is justified by image resolution and size, but no comparison with other base models is performed for the final pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AppVLM: A Lightweight Vision Language Model for Online App Control." pith.science (2026). https://pith.science/paper/IRKN7M54

@misc{pith2026250206395,
  author       = {Pith},
  title        = {Pith review of: AppVLM: A Lightweight Vision Language Model for Online App Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IRKN7M54}},
  note         = {Machine review of arXiv:2502.06395}
}
read the original abstract

The utilisation of foundation models as smartphone assistants, termed app agents, is a critical research challenge. These agents aim to execute human instructions on smartphones by interpreting textual instructions and performing actions via the device's interface. While promising, current approaches face significant limitations. Methods that use large proprietary models, such as GPT-4o, are computationally expensive, while those that use smaller fine-tuned models often lack adaptability to out-of-distribution tasks. In this work, we introduce AppVLM, a lightweight Vision-Language Model (VLM). First, we fine-tune it offline on the AndroidControl dataset. Then, we refine its policy by collecting data from the AndroidWorld environment and performing further training iterations. Our results indicate that AppVLM achieves the highest action prediction accuracy in offline evaluation on the AndroidControl dataset, compared to all evaluated baselines, and matches GPT-4o in online task completion success rate in the AndroidWorld environment, while being up to ten times faster. This makes AppVLM a practical and efficient solution for real-world deployment.

Figures

Figures reproduced from arXiv: 2502.06395 by the authors.

Figure 1
Figure 1. Visualisation of the RFT pipeline. Data is gathered by interactions between the emulators and AppVLM. The data is gathered, preprocessed and added to the dataset. It is used to perform a fine-tuning step. In this case, the parameters are the first name, surname, and phone number, allowing for a vast number of task varia￾tions. Our objective is to train an agent that can solve as many tasks as possible, using a light… view at source ↗
Figure 2
Figure 2. Example trajectory in AndroidWorld, with the goal at the top and the taken actions below each timestep’s screenshot.The agent almost succeeds in solving this task, but forgets to clear the text field before typing in the penultimate step. the agent almost completes the task correctly, but fails to clear the existing text before adding the filename in the penultimate step. This is a common mistake, where the agent do… view at source ↗
Figure 3
Figure 3. Example AndroidWorld observation passed as input to AppVLM. The visual input is composed of the current screenshot, annotated with bounding boxes surrounding clickable UI elements, along with numbered labels. The textual input is composed of the task goal, as well as the history of actions. This observation corresponds to the input for step 2 in [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Example trajectory in AndroidWorld, with the goal at the top and the taken actions below each timestep’s screenshot. AppVLM successfully creates an audio recording and saves it with the appropriate filename. Step 6 is noteworthy, with the agent opting for a long-press …
Figure 5
Figure 5. Figure 5: Example trajectory in AndroidWorld, with the goal at the top and the taken actions below each timestep’s screenshot. AppVLM successfully creates a new contact, filling out several form fields to do so. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Example trajectory in AndroidWorld, with the goal at the top and the taken actions below each timestep’s screenshot. AppVLM successfully sends a message to a specified phone number. open_app: Broccoli 0 click: 22 1 input_text: Shrimp Avocado Salad 2 click: 2 3 click: 1…
Figure 7
Figure 7. Figure 7: Example trajectory in AndroidWorld, with the goal at the top and the taken actions below each timestep’s screenshot. AppVLM successfully deletes a specific recipe, even when this recipe is not immediately visible in the list 14 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. MobileA3gent: Training Mobile GUI Agents Using Decentralized Self-Sourced Data from Diverse Users

    cs.AI 2025-02 conditional novelty 6.0 of 10

    A hierarchical auto-annotation pipeline plus episode-aware federated aggregation lets mobile GUI agents be trained on automatically labeled user trajectories at about 1% of human annotation cost.

  2. Software Engineering for and with GUI Agent

    cs.SE 2026-08 conditional novelty 5.0 of 10

    A survey of 336 GUI-agent papers finds rapid growth alongside weak engineering support for recovery, human oversight, maintainability, and privacy, and calls for lifecycle-centered testing and governance.

  3. Edge-Based Multimodal Sensor Data Fusion with Vision Language Models (VLMs) for Real-time Autonomous Vehicle Accident Avoidance

    cs.AI 2025-08 reject novelty 4.0 of 10

    A lightweight vision-language model on an edge device fuses roadside hazard alerts with onboard camera views to adjust trajectories, and the authors report a 77% simulated collision reduction over a vision-only baseline.

  4. Clip4Retrofit: Enabling Real-Time Image Labeling on Edge Devices via Cross-Architecture CLIP Distillation

    cs.CV 2025-05 conditional novelty 4.0 of 10

    Distilling CLIP into EfficientNet-B3 with projection heads yields a 24.6 MB model that performs zero-shot labeling on driving scenes in real time on an edge camera.

Reference graph

Works this paper leans on

21 extracted references · 1 canonical work pages · cited by 4 Pith papers

  1. [1]

    Digirl: Training in-the-wild device- control agents with autonomous reinforcement learning

    Bai, H., Zhou, Y ., Cemri, M., Pan, J., Suhr, A., Levine, S., and Kumar, A. Digirl: Training in-the-wild device- control agents with autonomous reinforcement learning. arXiv preprint arXiv:2406.11896,

    arXiv

  2. [4]

    Lightweight neural app control

    Christianos, F., Papoudakis, G., Coste, T., Hao, J., Wang, J., and Shao, K. Lightweight neural app control. arXiv preprint arXiv:2410.17883,

    arXiv

  3. [5]

    The llama 3 herd of models

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783,

    arXiv

  4. [6]

    Navigating the digital world as humans do: Universal visual grounding for gui agents

    Gou, B., Wang, R., Zheng, B., Xie, Y ., Chang, C., Shu, Y ., Sun, H., and Su, Y . Navigating the digital world as humans do: Universal visual grounding for gui agents. arXiv preprint arXiv:2410.05243,

    arXiv

  5. [7]

    L., Srinivasan, S., Konyushkova, K., Weerts, L., Sharma, A., Siddhant, A., Ahern, A., Wang, M., Gu, C., et al

    Gulcehre, C., Paine, T. L., Srinivasan, S., Konyushkova, K., Weerts, L., Sharma, A., Siddhant, A., Ahern, A., Wang, M., Gu, C., et al. Reinforced self-training (rest) for language modeling. arXiv preprint arXiv:2308.08998,

    arXiv

  6. [8]

    J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W

    Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685,

    arXiv

  7. [9]

    On the effects of data scale on computer control agents

    Li, W., Bishop, W., Li, A., Rawles, C., Campbell-Ajala, F., Tyamagundlu, D., and Riva, O. On the effects of data scale on computer control agents. arXiv preprint arXiv:2406.03679,

    arXiv

  8. [10]

    Infiguiagent: A multimodal generalist gui agent with native reasoning and reflection

    Liu, Y ., Li, P., Wei, Z., Xie, C., Hu, X., Xu, X., Zhang, S., Han, X., Yang, H., and Wu, F. Infiguiagent: A multimodal generalist gui agent with native reasoning and reflection. arXiv preprint arXiv:2501.04575,

    arXiv

Show all 21 references
  1. [13]

    J., Sodhani, S., Jayaraman, D., Bastani, O., Kumar, V ., and Zhang, A

    Ma, Y . J., Sodhani, S., Jayaraman, D., Bastani, O., Kumar, V ., and Zhang, A. Vip: Towards universal visual reward and representation via value-implicit pre-training. arXiv preprint arXiv:2210.00030,

    arXiv

  2. [15]

    Androidworld: A dynamic benchmark- ing environment for autonomous agents

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

    arXiv

  3. [16]

    Mmac-copilot: Multi-modal agent collaboration operating system copilot

    Song, Z., Li, Y ., Fang, M., Chen, Z., Shi, Z., Huang, Y ., and Chen, L. Mmac-copilot: Multi-modal agent collaboration operating system copilot. arXiv preprint arXiv:2404.18074,

    arXiv

  4. [17]

    Androidenv: A reinforcement learning platform for android

    Toyama, D., Hamel, P., Gergely, A., Comanici, G., Glaese, A., Ahmed, Z., Jackson, T., Mourad, S., and Precup, D. Androidenv: A reinforcement learning platform for android. arXiv preprint arXiv:2105.13231,

    arXiv

  5. [19]

    Ponder & press: Advancing visual gui agent towards general computer control

    Wang, Y ., Zhang, H., Tian, J., and Tang, Y . Ponder & press: Advancing visual gui agent towards general computer control. arXiv preprint arXiv:2412.01268, 2024d. Wen, H., Li, Y ., Liu, G., Zhao, S., Yu, T., Li, T. J.-J., Jiang, S., Liu, Y ., Zhang, Y ., and Liu, Y . Autodroid...

    arXiv

  6. [21]

    Gpt-4v (ision) is a generalist web agent, if grounded

    Zheng, B., Gou, B., Kil, J., Sun, H., and Su, Y . Gpt-4v (ision) is a generalist web agent, if grounded. arXiv preprint arXiv:2401.01614,

    arXiv

  7. [1992]

    Appagent: Multimodal agents as smartphone users

    Yang, Z., Liu, J., Han, Y ., Chen, X., Huang, Z., Fu, B., and Yu, G. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771,

    arXiv

  8. [2017]

    Comprehensive cognitive llm agent for smartphone gui automation

    Ma, X., Zhang, Z., and Zhao, H. Comprehensive cognitive llm agent for smartphone gui automation. arXiv preprint arXiv:2402.11941,

    arXiv

  9. [2021]

    Mobile-agent-v2: Mobile device operation assistant with effective nav- igation via multi-agent collaboration

    Wang, J., Xu, H., Jia, H., Zhang, X., Yan, M., Shen, W., Zhang, J., Huang, F., and Sang, J. Mobile-agent-v2: Mobile device operation assistant with effective nav- igation via multi-agent collaboration. arXiv preprint arXiv:2406.01014, 2024a. 9 AppVLM: A Lightweight Vision Lang...

    arXiv

  10. [2022]

    Agent q: Advanced reasoning and learning for autonomous ai agents

    Putta, P., Mills, E., Garg, N., Motwani, S., Finn, C., Garg, D., and Rafailov, R. Agent q: Advanced reasoning and learning for autonomous ai agents. arXiv preprint arXiv:2408.07199,

    arXiv

  11. [2023]

    Spa-bench: A comprehensive benchmark for smartphone agent evalu- ation

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

    2024

  12. [2024]

    S., Kolesnikov, A., Wang, X., Salz, D., Neumann, M., Alabdulmohsin, I., Tschan- nen, M., Bugliarello, E., et al

    Beyer, L., Steiner, A., Pinto, A. S., Kolesnikov, A., Wang, X., Salz, D., Neumann, M., Alabdulmohsin, I., Tschan- nen, M., Bugliarello, E., et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726,

    arXiv

  13. [2025]

    Fixing weight decay regu- larization in adam

    Loshchilov, I., Hutter, F., et al. Fixing weight decay regu- larization in adam. arXiv preprint arXiv:1711.05101, 5,

    arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.