acceptodds
Under review as a conference paper at ICLR 2027

MobileViews: Scalable and High-Fidelity Data Collection for Mobile GUI Agents

Abstract

Training capable mobile GUI agents requires diverse interaction data, yet collecting such data across a broad range of apps remains challenging. We introduce MobileViews, a mobile GUI dataset built through a scalable hardware–software pipeline combining native SoC clusters with VLM-assisted traversal and selective human fallback. It contains over 1.2M unique screenshot–view hierarchy pairs and associated interaction traces from more than 30K Android apps on Google Play. To assess the dataset’s utility, we construct MobileViewsIDM, comprising over 81K inverse dynamics samples from 26K apps for training models to understand GUI transitions. Across the three evaluated VLMs, training on MobileViewsIDM generally outperforms training on prior mobile GUI datasets that are limited in app diversity and scale. Our hardware–software co-design also improves native mobile app compatibility from 56% to 96% and increases the traversal success rate from 16.7% to 89.6%. MobileViews will be fully open-sourced.

Primary area:
datasets and benchmarks
Keywords:
mobile gui, agent, inverse dynamics, dataset

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GLF7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces MobileViews, a large-scale mobile GUI dataset collected from over 30K Android apps using SoC clusters, VLM-assisted traversal, and selective human intervention. The authors also construct MobileViewsIDM to evaluate the dataset for GUI transition understanding.

Critical strengths
  1. The scale and app diversity are strong. MobileViews contains over 1.2M screenshot–view hierarchy pairs from more than 30K apps and covers multiple screen resolutions.
  2. The downstream evaluation is reasonably broad. The authors evaluate three VLMs under multiple agent configurations and compare MobileViewsIDM against Rico and AndroidControl.
  3. The sample-size analysis in Figure 5 is useful because it shows that the gain is not explained only by having more training samples.
Critical weaknesses
  1. Figure 2(b) presents a relatively simple UI transition graph, but real mobile interaction graphs can be much more cyclic due to back/navigation actions and varying starting states. It would be helpful to clarify how the actual UTG is constructed and whether this figure is only a simplified illustration.
  2. Table 2 is intended to demonstrate the value of MobileViews as training data, rather than model superiority. The sample-size analysis in Figure 5 is useful, but the comparison still mixes dataset size, app diversity, and data quality. Can the authors better isolate which factor contributes most to the gain?
  3. The downstream evaluation mainly focuses on inverse dynamics prediction. While this is a reasonable use of the collected trajectories, it is still only one capability.
Questions for authors
  1. Are system UI elements such as time, battery, signal, and navigation bars retained in the raw screenshots? If so, are they considered in the dataset diversity analysis?
  2. How much human effort is required by the fallback stage, e.g., interventions per app or operator time per successful traversal?
3 comments
AuthorsOpenReview ↗

Response to Reviewer GLF7

We thank the reivewer for the thoughtful reviewers. We respond to the concerns as follows.

W1: Distinguishing between an actual UTG and the demonstrated one in Figure 2(b)

Figure 2(b) shows a simplified version of the complete UTG we collected for the “Mood Tracker” app. The complete UTG contains 213 unique GUI states and 312 directed edges. As the reviewer said, an app’s complete UTG contains more cycles than the figure shows because both forward actions (e.g., click and scroll) and backward actions (e.g., navigate_back) can lead to previously visited states.

W2: Contribution to model performance gains

We agree with the reviewer that Table 2 shows the overall benefit of MobileViewsIDM, yet Figure 5 shows that MobileViewsIDM can achieve higher IDM accuracy with fewer training samples compared to other datasets. This suggests that training-set size alone does not explain the gains. Broader app coverage is one plausible contributor. For example, PhoneWorld [1] found that training on more apps improved AndroidWorld performance, while ProCUA-SFT [2] found that balancing training trajectories across app combinations raised performance on OSWorld. In the next revision, we will analyze how IDM accuracy changes with the number of distinct training apps while holding the sample count fixed.

[1] Liu, et al. “PhoneWorld: From Real-App Trajectories to Dynamic and Verifiable Environments for Phone-Use Agents.” arXiv preprint arXiv:2605.29486, 2026. [2] Jung, et al. “ProCUA-SFT Technical Report.” arXiv preprint arXiv:2606.17321, 2026.

W3: IDM is only one evaluated capability

We have additionally evaluated the trained checkpoints of different data scales on GUI interaction tasks. Please refer to Common Response #3.

Q1: Details on dataset diversity analysis

The collected GUI screenshots are as-is in MobileViews, without removing system UI elements. For the dataset diversity analysis, we compute a 64-bit perceptual hash for each full screenshot and treat two screenshots as visually similar if the Hamming distance between their hashes is at most 5. Under this criterion, screenshots that differ only in their system UI bars are typically classified as visually similar. We selected this heuristic threshold after a separate check performed before deduplication, in which 3,197 of 3,198 same-app screenshot pairs that differed only in the status bar met the similarity criterion.

Q2: Human intervention effort

Please refer to Common Response #2.

We hope the above responses could address the reviewer’s concern, and we are willing to talk about any further feedback.

Thank you for the response. The UTG and system-bar clarifications resolve W1 and Q1, and the AndroidControl results are a useful addition. However, my main concern (W2) remains open: the analysis isolating app diversity from sample count is deferred to a future revision. Similarly, the AndroidControl evaluation shows that IDM training helps, but without Rico/AndroidControl-trained baselines it does not show that MobileViews is a better training source, and the gain shrinks to ~1.5pt at full scale. I appreciate the effort and continue to view the paper positively, so I will keep my score at 3.

Thank you for taking the time to read our response and follow up. We sincerely appreciate your positive assessment of MobileViews, especially your recognition of its scale and app diversity. Your questions about the UTG, GUI deduplication, and human intervention have also helped us improve the clarity of our paper. We are glad that our clarifications addressed your earlier concerns about the UTG and system bars.

We also appreciate your suggestions for further evaluation. We will do our best to include an analysis of app diversity at matched sample counts and additional AndroidControl results from checkpoints trained on different datasets in the next revision. We hope to further explore the potential of MobileViews as an open-source resource for mobile GUI research. Thank you again for your time, positive feedback, and continued support of our work.

Reviewer_DQA7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MobileViews, a large-scale data collection framework designed for mobile GUI agents. By combining a cluster of native mobile SoCs, rule-based interface traversal, assistance from vision-language models (VLMs), and human intervention, the framework constructs the MobileViews dataset—comprising over 1.2 million pairs of screenshots and view-hierarchy (VH) data, spanning more than 30,000 Android applications. The authors further develop "MobileViewsIDM" through a multi-stage filtering process and conduct inverse dynamics modeling (IDM) experiments using three different VLMs to evaluate how the dataset enhances the understanding of GUI state transitions.

Critical strengths
  1. The scale of the data and the breadth of application coverage are notable; MobileViews includes over 1.2 million screenshot-VH pairs covering more than 30,000 Android apps.
  2. The authors compare various training data sources across three VLMs and three agent configurations, providing empirical evidence regarding the dataset's utility for GUI inverse dynamics tasks.
Critical weaknesses
  1. While the paper highlights scalability and efficient data collection as key selling points, the reported experiments focus primarily on application compatibility and traversal success rates, failing to adequately quantify the actual resources required to achieve these results.
  2. The high traversal success rate is heavily dependent on human intervention, leading to an overestimation of the framework's level of automation.
  3. The multi-stage VLM-based filtering process may introduce selection bias, and there is a lack of independent verification regarding data quality.
  4. Although the paper reports IDM accuracy across various models and agent configurations, it fails to provide the variance associated with multiple independent training runs.
Questions for authors
  1. The methodological innovation is primarily centered on the system integration of existing components, lacking a sufficiently distinct, independent technical contribution.
  2. The system's scalability and cost-effectiveness are not rigorously quantified, and the final traversal success rate remains significantly reliant on human intervention.
3 comments
AuthorsOpenReview ↗

Response to Reviewer DQA7

We thank the reviewer. Regarding the concerns, we made the following clarifications.

W1: Resources to collect the dataset

Please refer to Common Response #2.

W2/Q2: Dataset collection requires a lot of human intervention

Please refer to Common Response #1. In summary, based on the new experiment using GPT-6 Astra as the backend VLM, we believe that many of the cases that required human involvement in our micro-benchmarks on the top 100 Google Play apps can be resolved, while human intervention is still necessary for more comprehensive and diverse data collection.

W3: Data quality verification

We understand the reviewer’s concern about the selection bias incurred by the VLM-based filtering process. Here, we conduct human validation on 150 samples retained in the final dataset and 50 samples rejected at each of the three filtering stages. Three co-authors of this study independently review each sample without knowing whether it was retained or rejected. They first infer the action and its arguments from the before-and-after screenshots and note any plausible alternatives. We then show them the recorded action and ask whether the before-and-after screenshots provide enough evidence to distinguish that action and its arguments from any plausible alternatives. We record agreement with the VLM’s filtering decision when at least 2 of the 3 human validators judge a retained sample usable or a rejected sample unsuitable. The following table shows the results.

Human reviewers judged 139 of the 150 retained samples (92.7%) usable. Disagreements included cases where two different actions could produce the same screen transition. For rejected samples, reviewers agreed with the filtering decisions in 100%, 92%, and 88% of cases at Stages 1/2/3, respectively. Exemplified cases include: At Stage 2, the filter may reject a transition because the next screen is only partially rendered or the entered text is synthetic, even when reviewers can identify the action and its arguments from the screenshots. At Stage 3, the model’s predicted click may fall outside the annotated bounding box even when the screenshots clearly indicate which element was tapped. Overall, most retained samples are suitable for IDM, although the later filters also discard a few usable samples.

Data sample sourceSelected samplesHuman-VLM agreement samples
Retained samples in the final MobileViewsIDM150139
Rejected samples at Stage 15050
Rejected samples at Stage 25046
Rejected samples at Stage 35044

We hope this human validation addresses some of the reviewer’s concerns. Given the limited rebuttal period, we evaluated only a small sample. We will expand the human evaluation and report the results in the revised manuscript to provide a more precise estimate of MobileViewsIDM’s quality. We will also fully release MobileViews and MobileViewsIDM so that researchers can independently assess their quality and construct additional high-quality samples from the original MobileViews.

W4: Multiple independent training runs

We appreciate the reviewer’s concern. Yet conducting multiple independent training runs using GRPO on around 175K samples is computationally demanding and exceeds our available GPU budget. We will note this limitation in the paper and leave measuring run-to-run variance to future work.

Q1: Technique contribution beyond “integrating existing system components”

We position this study within the “datasets and benchmarks” track of ICLR. We clarify the contributions of this study as follows.

  • We anticipate that one of the core contributions lies in the large-scale, diverse mobile GUI dataset, which we will open-source for future usage.

  • Beyond the dataset, we observe the challenges of collecting a diverse, in-the-wild dataset and propose a hardware–software design for scalable data collection. Section 5.3 evaluates how these design choices improve app compatibility and traversal success, which, to our knowledge, has not been covered by prior studies.

  • We present a benchmark, MobileViewsIDM, for training and evaluating GUI transition understanding.

As for the system aspect, prior data collection approaches are limited by app compatibility on x86 emulators, by (1) rule-based traversal that can get stuck in complex GUI states, and by (2) costly human annotators. Our hardware-software co-design addresses both challenges, enabling us to collect screenshot-VH pairs and interaction trajectories at scale across diverse real-world Android apps.

We hope the above responses could address the reviewer’s concern, and we are willing to talk about any further feedback.

Thank you to the author for the response, most of my questions have been addressed, so I am considering raising my rating.

We thank the reviewer for your time and constructive feedback throughout the review process. We are glad our response helped address most of the concerns. We would be happy to discuss any remaining questions you may have.

Reviewer_mXoNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MobileViews, a large-scale dataset for mobile GUI agents, comprising 1.2 million Screenshot–View Hierarchy data pairs and interaction trajectories collected from over 30,000 Android apps. By combining SoC clusters, rule-based traversal, VLM assistance, and human intervention, the authors implement a large-scale GUI data collection pipeline. Furthermore, the authors construct MobileViewsIDM and validate the dataset's value through IDM training experiments using three VLMs under various configurations, while also evaluating the compatibility and traversal success rate of the collection system.

Critical strengths
  1. Large-scale and comprehensive data resources. MobileViews exhibits distinct advantages in app coverage, data scale, and modal completeness. Compared to existing datasets, it simultaneously provides screenshots, view hierarchies, and interaction trails, offering a valuable data resource for GUI understanding and interactive learning.

  2. A practical data collection pipeline. The authors combine SoC clusters with semi-automated traversal mechanisms to improve app runtime compatibility and data collection success rates, as demonstrated by the clear system effects in Tables 3 and 4. Additionally, IDM training experiments across multiple models preliminarily prove the practical value of the new dataset.

Critical weaknesses

W1. The paper claims high scalability for its data collection framework. However, Table 4 shows that the automated traversal success rate (Rule-based + VLM) is only 39.6%, jumping to 89.6% only after human intervention.

W2. The data quality is evaluated solely on the proposed MobileViewsIDM benchmark. Even with an app-disjoint split, models trained on MobileViewsIDM share the same platform-specific collection and filtering distribution during test time, giving them an inherent advantage over baseline datasets (e.g., Rico, AndroidControl).

W3. Unexplained performance anomalies and lack of variance analysis. In Figure 5, performance scales non-monotonically for certain models as training data increases (e.g., Qwen3-VL-8B-Instruct under the SoM configuration). Additionally, Table 2 shows that MobileViewsIDM slightly underperforms Rico for Qwen3-VL-8B-Instruct under None-CoT and CoT settings. The paper provides no substantive analysis to explain these performance fluctuations.

Questions for authors

Q1. Table 4 shows that human intervention has a significant impact on the final traversal success rate. Can the authors provide the proportion of apps requiring human handling, average intervention count, human work hours, and unit effective data collection cost throughout the entire data collection process?

Q2. Section 3.2 mentions using Droidbot to sequentially execute interactions with a maximum budget of 1,000 actions. Could the authors further clarify whether BFS, DFS, or other specific traversal strategies are adopted? Are visited GUI states and state–action pairs maintained to avoid redundant exploration? Additionally, can they report the number of unique GUI states per app, the proportion of duplicate interactions, and how the number of new GUI states changes as the action budget increases?

Q3. In Figure 5, the performance of certain configurations drops locally as training samples increase (e.g., the SoM configuration of Qwen3-VL-8B-Instruct). Furthermore, the improvements of MobileViewsIDM on stronger models in Table 2 are unstable. Can the authors explain the sample selection approach across different training scales to determine whether these variations are statistically significant, and what causes different models to benefit from MobileViewsIDM to varying degrees?

Q4. External Generalization: Can the authors provide external evaluation results independent of the MobileViewsIDM construction pipeline to verify the generalization value of MobileViews as training data for other GUI understanding or interaction tasks?

1 comment
AuthorsOpenReview ↗

Response to Reviewer mXoN

We thank the reviewer for the constructive reviews. We respond to the concerns in the following.

W1/Q1: High traversal success rate only after human intervention, and quantization on human intervention

We thank the reviewer for raising the crucial concern. Please refer to Common Response #1 and #2 for a detailed clarification.

W2/Q4: Evaluation beyond MobileViewsIDM

Please refer to Common Response #3. We have incorporated the new experimental results into the revised manuscript and will further clarify how models trained on IDM tasks may generalize to GUI interaction tasks.

W3: Training data sampling strategy & performance fluctuation on Qwen3-VL-8B-Instruct under different agent setups

  • Training data sampling strategy: We randomly sampled MobileViewsIDM training examples from the apps in the training split, with each app contributing at most five GUI transition samples, as described in the data filtering process.

  • Performance fluctuations in Figure 5: Before fine-tuning, Qwen3-8B already achieves 74.04% IDM accuracy under SoM, compared with 29.90% under None-CoT and 31.84% under CoT. This indicates that the model can already use SoM’s explicit element labels to infer action targets, leaving less room for improvement through fine-tuning. The narrow y-axis range in the Qwen3-8B panel of Figure 5(c) makes the local dip visually prominent. In absolute terms, accuracy decreases by 0.23% from 10K to 20K examples. Since performance continued to improve afterward, we consider this drop within the normal range of training variation.

  • MobileViewsIDM underperforms Rico under two agent setups in Table 2: We observe that models trained on the two datasets show near-identical performance under these two setups, with accuracy differences of only 0.14% and 0.47%. Yet models trained on MobileViewsIDM exhibit higher performance under most setups shown in Table 2 and Figure 5 (7 of 9 model-configuration pairs). Taken together, we believe these results do not undermine the core claim of this study.

Q2: Traversal algorithm, GUI states vs. interaction budgets

We use DroidBot’s default depth-first search (DFS) to explore apps. To reduce redundant exploration, it maintains a UI transition graph of visited GUI states and records which actions it has tried in each state. When it encounters a state again, it prioritizes untried interactions. Once it has exhausted the actions in the current state, it can replay known transitions to reach another state with unexplored actions. This enables it to revisit a screen for navigation without treating that screen’s previously tried actions as new exploration opportunities. The unique GUI states per app are demonstrated in Figure 3(a) in this paper, with more than 100 unique GUI states in each of 1,168 apps.

To quantify “the proportion of duplicated interactions, and how the number of GUI states changes as the action budget increases”, we analyzed DroidBot’s raw interaction log. We have added a figure of the results to the revised manuscript (Figure 6 on Page 20). The average number of unique GUI states per app rises rapidly during the first few interactions because most actions reach previously unseen states. As exploration continues, actions increasingly revisit previously seen states, and the average grows more slowly, reaching 33.54 states at 250 interactions, 38.66 at 500, 39.89 at 750, and 39.99 at 1,000. Therefore, the number of duplicate interactions is low in the first few steps and increases substantially in later steps.

We hope the above responses could address the reviewer’s concern, and we are willing to discuss any further feedback.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Common Response (1/2)

We thank all reviewers for their valuable feedback. We appreciate that all three reviewers highlighted the scale and app coverage of MobileViews, which spans more than 30,000 Android apps and includes over 1.2 million screenshot–view hierarchy pairs along with interaction trajectories. Reviewers also noted that our comparisons with Rico and AndroidControl across three VLMs and multiple configurations show the value of MobileViews as training data for GUI inverse dynamics prediction.

For concerns raised by the reviewers, we have included additional evaluation results and highlighted the revisions in color in the updated manuscript. We first respond to common concerns and then respond to each reviewer individually.

Common Response #1: Data collection automation level: A non-trivial portion of apps require human intervention in micro-benchmarks

We thank the reviewers for raising the concern about the automation level of the dataset collection pipeline. The current experiment on app traversal success rate on the top-100 Google Play Store apps demonstrates that a fully automated approach only solves around 40% of all apps as of the date when we conducted the experiment, and leaves a non-trivial portion of apps requiring human intervention during app exploration. We clarify such a result from three perspectives.

  1. Popular apps can be more difficult to traverse automatically because they often use CAPTCHAs and other anti-automation measures; they may restrict access from emulators, or require personal information for verification. These steps often require human intervention, after which rule-based traversal can resume and explore the remaining GUI states. In our experience, less popular apps require substantially less human intervention than the popular apps evaluated in Table 4.

  2. At the time of the experiment, we selected GPT-4o as the backend VLM for its cost efficiency. We used the same model for both large-scale data collection and the evaluation of the top 100 apps so that the results for those apps would better reflect performance across the full app set. Although GPT-4o is cost-efficient, its capabilities fall behind nowaday SOTA VLMs such as GPT-6 Astra and Claude Fable 5.2. We expect that these models could handle the observed errors more effectively and reduce the need for human intervention.

  3. As our approach requires human intervention, we described it as “semi-automated” in the original manuscript to avoid overstating its level of automation.

To further support our expectation, we conducted additional experiments in which we replaced GPT-4o with GPT-6 Astra, a more recent SOTA VLM to date, to test whether it could resolve prior unsuccessful cases. Before conducting the experiments, we examined the 48 cases that required human intervention for a successful traversal, and classified them by the apparent cause of failure. The table below summarizes this classification.

Empirical attributionGPT-4o successGPT‑6 Astra successExemplified failed apps/cases in the Experimental setup
Limited VLM capability30/4824/30happn missed a visible Next button. Grindr stalled at an image CAPTCHA. Boo and Tappytoon required switching to an inbox to confirm signup.
Unavailable/failed UI capture/input10/486/10Marvel returned neither a screenshot nor a UI hierarchy. Telegram rejected the crawler’s text input.
Unmet access requirements or app errors8/482/8WhatsApp rejected the test number. cloudLibrary required a library ID. Revolut Business returned an error even during manual signup.

We attribute that most of the prior failed apps/cases (30 out of 48, 62.5%) are due to limited VLM capability, i.e., GPT-4o cannot correctly locate the target GUI element, or fails to pass the CAPTCHA. With GPT-6 Astra as the backend, our pipeline successfully traverses 32 of the 48 previously failed cases in total (due to the constrained rebuttal period, we only set a 150-step threshold). We therefore expect stronger VLMs to increase the proportion of apps our pipeline can traverse autonomously. We retain selective human intervention for in-the-wild data collection because some steps require access or decisions beyond the agent’s available tools, such as retrieving a verification code sent to a user’s phone or deciding whether to provide personal information.

To summarize, we hope this additional experiment addresses the reviewers’ concern by showing how advances in VLM capabilities can improve the automation level of the data collection approach. Human intervention remains necessary to cover a broader range of apps and collect more diverse data, which aligns with our original claim that the approach is semi-automated.

AuthorsOpenReview ↗

Common Response (2/2)

Common Response #2: More statistics on efficiency and costs of dataset collection

We thank the reviewers for raising this point that would help the readers better understand the efficiency of dataset collection process. We collected the dataset primarily on two SoC clusters that together support 240 virtualized Android instances. Collection took approximately 151,600 device-hours in total.

For monetary costs, we used GPT-4o as the backend VLM. Our pipeline first attempts rule-based traversal, invokes the VLM when traversal is blocked, and requests human intervention if the blocker persists. During collection, the pipeline typically invoked the VLM to handle registration/signin in the first few steps or to dismiss unprocessed pop-ups by the VLM alone. The pipeline averaged 1.75 VLM calls per app. We observed that most calls occurred in widely used apps with more GUI states. The estimated cost of VLM calls across all ~30K apps was $524.

Of the 30,352 apps, 9,138 (30.1%) required human intervention, with 11,057 interventions in total. Four co-authors handled these cases in batches. We need to clarify we did not record their working hours during dataset collection, but in most apps, a human could resume traversal within a few seconds for handling popups. Cases requiring registration or form completion took longer, ranging from 3 to >10 minutes. We have updated the manuscript to include the device-hours, monetary cost, and human effort to clarify the resources required for data collection.

Common Response #3: Evaluation on more benchmarks beyond MobileViewsIDM

Our initial consideration of evaluating models only atop MobileViewsIDM is that the prior study, GUI-Shift [1], had already shown that models trained on inverse dynamics examples can generalize to related tasks, including GUI interaction tasks [2, 3].

Here, to further show the utility of our collected dataset, we also evaluate the Qwen3-VL-4B and Qwen3-VL-8B checkpoints trained on MobileViewsIDM on AndroidControl, which includes 10,814 static GUI interaction tasks. The results are shown in the following table. The core observation is that, both models trained on MobileViewsIDM exhibit non-trivial performance gains on GUI interaction benchmarks. Yet, these gains decrease as the number of training samples increases.

We hypothesize that with more IDM training, the model becomes more specialized for inferring past actions from differences between consecutive screens. This may explain the smaller gains on AndroidControl, where the model must predict the next action from the current screen and the instruction alone. Mixing IDM and instruction-following examples may help preserve these gains at larger training scales.

# training samplesQwen3-VL-4B-InstructQwen3-VL-8B-Instruct
0 (base model)79.6184.73
1K86.22 (+6.61)90.01 (+5.28)
2K86.30 (+6.69)89.95 (+5.22)
3K86.23 (+6.62)90.06 (+5.33)
4K86.22 (+6.61)90.05 (+5.32)
5K86.07 (+6.46)90.02 (+5.29)
10K81.59 (+1.98)86.24 (+1.51)
30K81.51 (+1.90)86.26 (+1.53)
50K81.28 (+1.67)86.04 (+1.31)
Full (~65K)81.08 (+1.47)86.19 (+1.46)

We have included experimental results in Sec. 5.3, Page 9, to enrich the manuscript. We will further clarify the generalization potential when training on the selected IDM manuscript. We welcome any further feedback from the reviewers.

[1] Gao, et al. "GUI-Shift: Enhancing VLM-Based GUI Agents through Self-Supervised Reinforcement Learning." International Conference on Learning Representations, 2026. [2] Li, et al. "On the Effects of Data Scale on UI Control Agents." Advances in Neural Information Processing Systems 37, 2024. [3] Rawles, et al. "AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents." International Conference on Learning Representations, 2025.

Area_Chair_41CyOpenReview ↗

Reminder: Please Review Author Rebuttals and Participate in Discussions

Dear Reviewers,

Thank you for your time on the reviews. The authors have now submitted their rebuttals, and I encourage you to read their responses and participate in the discussion.

Please consider whether the rebuttals address your concerns and, where clarification is still needed, engage with the authors through the discussion forum. I also encourage you to exchange perspectives with your fellow reviewers and update your review or rating if appropriate.

Your participation will help ensure a thorough and fair evaluation of each paper. Thank you again for your time and effort.

Best regards,

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.