acceptodds
Under review as a conference paper at ICLR 2027

Imagine to Act: High-Fidelity Data Synthesis via Image Editing World Model for Scalable GUI Agent Training

Abstract

Graphical User Interface (GUI) agents have emerged as a promising paradigm for automating complex digital workflows across diverse applications. However, training highly capable and generalizable agents fundamentally relies on massive, high-fidelity visual-action trajectories, which are notoriously difficult to acquire. While human demonstrations are unscalable, existing GUI world models rely on text descriptions or HTML rendering, discarding crucial pixel-level visual details like icons and layout styles. To address this issue, we introduce Infinite-Dreamer, a simulation-free data synthesis method powered by a pixel-level Image Editing World Model. By conceptualizing GUI transitions as image editing tasks, we leverage Vision-Language Models (VLMs) to describe action-induced UI changes as structured delta-text. We then fine-tune an image editing backbone to controllably synthesize realistic screenshot transitions. We utilize this model to generate both single-frame visual robustness data and multi-step imaginary trajectories. To validate the effectiveness of our approach, we fine-tune the Qwen3-VL baseline solely on the synthesized data to obtain Infinite-Actor, and evaluate it on AndroidWorld, MobileWorld, and AndroidControl-Curated benchmarks. Infinite-Actor consistently outperforms the Qwen3-VL baselines across scales: Infinite-Actor-8B improves AndroidWorld Pass@1 by +4.45 and nearly doubles the MobileWorld Pass@3 success rate, while Infinite-Actor-2B improves Pass@1 by +9.05.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
GUI Agent, Data Synthesis, World Model, Image Editing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qsi4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Infinite-Dreamer, a synthetic data generation framework for training mobile GUI agents. It models GUI transitions as image-editing operations, using structured delta-text to generate high-fidelity next-screen images and multi-step imagined trajectories. The synthesized data is then used to fine-tune Qwen3-VL into Infinite-Actor.

Critical strengths

The image-editing world model produces relatively high-quality GUI transitions, preserving layout, text, and other visual details better than the compared general-purpose image-editing baselines, including in multi-step rollouts.

Critical weaknesses

1.The paper motivates synthetic data as a more scalable alternative to real trajectory collection, but does not provide an end-to-end cost comparison. Given the relatively high inference cost of the image-editing world model, it is unclear whether generating and filtering synthetic trajectories is actually cheaper than collecting real interactions at comparable scale.

2.Infinite-Actor improves task success on AndroidWorld and MobileWorld, but the results on AndroidControl-Curated are mixed: Step SR decreases in three of the six settings, and coordinate grounding accuracy drops in most settings. The paper also does not compare synthetic and real data under the same training-data budget, making it difficult to determine whether the proposed synthetic data is more effective or data-efficient than real trajectories.

Questions for authors

Could the authors compare synthetic and real trajectories under the same training-data budget? This would make it easier to assess the actual data efficiency of the proposed synthesis pipeline.

2 comments

Thank you for recognizing our contribution. We hope the following reply can address your concern.

1. Response to Q1: Budget Comparison

We conduct matched-budget comparisons at two levels. We first extend the data-volume ablation to 27,360 samples, with a 9:1 real-to-synthetic ratio and matched optimization budget and task coverage. Under this control, adding synthetic data improves AndroidWorld success to 44.83%, and Real+Aug+Gen reaches 48.28%:

ArmRealSyntheticPoolAW SR
Real only27,360027,36042.67
Real + Gen24,6242,73627,36044.83
Real + Aug24,6242,73627,36047.41
Real + Aug + Gen24,6242,73627,36048.28
Pure synthetic027,36027,36039.37

We further compare real and synthetic data with Qwen3.5-4B under matched budgets. Synthetic data scales better, surpassing real data at a larger scale with lower cost:

Data budget0(baseline)1k2k3k6k
Real only37.5035.3443.5043.97-
Synthetic only37.5032.7639.2240.0945.69

we have revised our manuscript accordingly. (see PDF page 8, Section 4.4 and Table 3; PDF page 46, Appendix I.1; PDF page 48, Appendix I.3)

2. Response to W1: Cost Comparison

Our cost analysis shows that our pipeline is more cost-efficient, achieving higher downstream performance than real-data collection at a much lower total cost when scaled up. It costs only $0.032 per step, 9.6×, 11.9×, and 10.8× cheaper than OpenMobile, AndroidControl, and AITZ, respectively. 6k synthetic samples reach 45.69% AndroidWorld success at $196.61, versus 3k real samples at 43.97% and $940.03.

We also provide cost breakdowns against OpenMobile and Code2World, plus the costs of public datasets. The one-time world-model fine-tuning cost is $432 and does not scale with data volume, so our pipeline becomes more cost-effective as the dataset grows.

OpenMobileCostCode2WorldCostOursCost
emulator$0.1316world model$0.0205Diffusion$0.0290
Monitoring$0.1629Action generation$0.0071Quality filtering$0.0013
Weak Model$0.0110Quality filtering$0.0029App Designer$0.0008
————text-only delta$0.0010
Total Per Step$0.306/0.066hTotal Per Step$0.0306/0.057hTotal Per Step$0.032/0.019h
DatasetAndroidControl[1]AITZ[1]
Cost$0.377$0.345

We further fine-tune Qwen3.5-4B under different generation and collection budgets. Although real data is better at the same volume, ours performs better at a larger scale with substantially lower total cost:

MethodDataBudgetTotalCostAndroidWorld(SRPASS@1)
OpenMobile3k$940.03/202.75h43.97
Code2World2k$69.9/116.05h37.07
Ours6k$196.61/116.74h45.69

we have revised our manuscript accordingly. (see PDF pages 47-48, Appendix I.2; PDF page 48, Appendix I.3)

[1] Rui Lv, et al. "M$^2$-Miner: Multi-Agent Enhanced MCTS for Mobile GUI Agent Data Mining" ICLR 2026

3. Response to W2: AC-Curated Regression

We add a real-4B control under the same protocol, comparing base-4B, real-4B, and Infinite-Actor-4B on AC-Curated-Hard. SFT and synthetic data contribute comparably to the Box Acc drop, while synthetic data contributes more to the Step SR drop; the regression thus stems partly from synthetic data and partly from fine-tuning:

ModelBox AccStep SR
base-4B77.0057.10
real-4B71.0352.31
Infinite-Actor-4B64.7044.20

We further analyze the coordinate errors: the Hard median center distance grows from 0.18 to 0.38 annotated-box widths, 97% of incorrect boxes place their center more than 50 px outside the annotation, and Box Acc is 46.7% on the smallest third of targets versus 75.4% on the largest third. The loss is systematic: most errors are far outside the annotation and concentrate on compact targets, explaining why the offline metric drops while online task success improves at all three scales:

DifficultyMetricBaseInfinite-Actor
EasyMedian center distance0.060.13
HardMedian center distance0.180.38
EasyWrong boxes>50 px outside—88%
HardWrong boxes>50 px outside—97%
HardBox Acc, smallest / largest third67.5/87.046.7/75.4

we have revised our manuscript accordingly. (see PDF page 8, Section 4.3; PDF page 44, Appendix G.3; PDF pages 48-49, Appendix I.5)

Thank you very much for your comment. We have supplemented relevant experiments to compare the gap between synthetic data and real data, and reported the required costs:

  1. Synthetic data is scalable, cost-effective, and complementary to real data Cost Report show that our pipeline costs only $0.032 per step, 9.6–11.9× less than OpenMobile and manual collection from AndroidControl and AITZ and supplementary experiments show that synthetic data can achieve a higher success rate than real data at substantially lower cost(45.69, $196.61 vs 43.97, $940.03), demonstrating its scalability and cost-effectiveness. Moreover, under the same data budget, adding synthetic data improves AndroidWorld success to 44.83%, and Real+Aug+Gen reaches 48.28%, highlighting the complementary value of synthetic data.

  2. The AC-Curated Regression stems from fine-tuning Our experiments show that SFT and synthetic data contribute comparably to the Box Acc drop. The regression therefore stems partly from fine-tuning and partly from synthetic data . Further analysis of coordinate errors indicates that the loss is systematic: most errors lie far outside the annotation and concentrate on compact targets, which explains why the offline metric drops while online task success improves at all three scales.

  3. Our core innovation is a unified state representation Our core innovation is a unified state representation that simultaneously accounts for visual realism and operational feasibility, something existing GUI world models cannot achieve. we represent GUI transitions as structured, coordinate-free deltas encoding trigger operations, added/removed elements, and subsequent affordances. This representation enables both learning from real interaction pairs and generation from text-only plans, bridging the gap between visually faithful simulation and executable GUI interactions. None of prior GUI world model has achieved such goal before.

We hope this response has addressed your questions. Please feel free to raise any further issues at any time; we are happy to continue the discussion.

Reviewer_KoVBOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Infinite-Dreamer, a simulator-free framework for synthesizing training data for mobile GUI agents. It represents GUI transitions using structured delta-text that describes the triggering operation, added and removed UI elements, and optional next-step affordances. A pretrained image-editing model is fine-tuned on screenshot pairs and delta-text to generate subsequent GUI states. The framework supports both appearance-preserving single-frame augmentation and multi-step imagined trajectories, with VLM-based filtering to remove low-quality samples.

Critical strengths

1.Fine-tuning on the synthesized data yields modest performance improvements over the Qwen3-VL baseline in some settings. 2.The trajectory synthesis pipeline offers potentially useful insights for constructing GUI training data.

Critical weaknesses
  1. The methodological novelty is insufficiently established. The framework combines structured UI-change descriptions, image-editing fine-tuning, synthetic trajectory generation, and VLM-based filtering. Although this combination is useful, the paper should more clearly identify what is technically new relative to existing GUI world models and data synthesis pipelines. Infinite-Actor itself is a fine-tuned Qwen3-VL model and therefore does not constitute an independent methodological contribution.

2.The experiments do not sufficiently isolate the value of the proposed synthesis method. Comparisons with the base Qwen3-VL models demonstrate that additional fine-tuning helps, but do not establish whether this particular synthesis pipeline is better than alternative sources of supervision. Table 3 includes a useful real-data baseline, yet comparisons with matched training budgets and alternative synthesis methods are needed to distinguish improvements due to data quantity, appearance augmentation, and the proposed multi-step imagination mechanism.

3.The gains are inconsistent across evaluation settings. Although online task success improves, coordinate grounding deteriorates in several configurations. For example, on AC-Curated-Hard, the 4B model’s Box Accuracy drops from 77.00% to 64.70%, and Step Success Rate falls from 57.10% to 44.20%. Moreover, mixed-data training reduces success on the hardest AndroidWorld tasks from 13.2% to 10.5%. These regressions limit the claim that synthetic supervision consistently improves agent generalization.

4.The incremental benefit of multi-step imagination remains unclear. In Table 3, adding appearance augmentation to real data improves AndroidWorld success from 42.67% to 47.41%, whereas adding imagined trajectories improves it to 44.83%. Adding imagined trajectories on top of real data and augmentation yields only a further 0.87 percentage-point gain. Without uncertainty estimates for these comparisons, the evidence for the additional value of the more complex trajectory-generation component remains limited.

5.Long-horizon synthesis remains a substantial limitation. The reported world-model quality decreases with rollout length, and only 9.3% of trajectories reaching step 10 pass filtering at every preceding step. This raises concerns about the efficiency of producing usable long trajectories and whether the retained data adequately covers difficult workflows.

6.Visual fidelity does not establish executable transition correctness. The image editor receives detailed delta-text describing the intended UI changes, while the App Designer specifies the imagined workflow. Consequently, convincing screenshots alone do not demonstrate that the generated transitions follow the behavior of a real application. Stronger validation of action–state consistency and task completion would help substantiate the world-model claim.

Questions for authors

1.Could you expand Table 1 to include more recent, competitive models from the Claude, Gemini, and GPT families? Please specify the exact versions and evaluate them under the same protocol to better contextualize Infinite-Actor’s performance.

2.Could you evaluate the proposed synthetic data on additional backbones, such as Qwen2.5-VL and the Qwen3.5 series? The current experiments cover multiple model sizes within Qwen3-VL, but do not establish whether the benefits generalize across model families or generations.

4 comments

Thank you for recognizing the contributions and experimental section of our work. We hope the following responses address your concerns and questions.

1. Response to Q2: Transferability to Qwen3.5

Our below results demonstrate that the proposed synthesis pipeline generalizes beyond Qwen3-VL to Qwen3.5-4B, confirming its cross-generation transferability. At 6k samples, synthetic data improves AndroidWorld success from 37.50% to 45.69% (+8.19 points), surpassing the best real-data result (43.97% at 3k).

Data budget0(baseline)1k2k3k6k
Real only37.5035.3443.5043.97-
Synthetic only37.5032.7639.2240.0945.69

The transfer is enabled by our model-agnostic (s, g, a) trajectory representation, which is compatible with the shared GUI action space rather than model-specific features.

we have revised our manuscript accordingly. (see PDF page 48, Appendix I.3)

2. Response to Q1: Competitive Models

We evaluate three recent models under the same protocol. Results show that all of them complete the 116 AndroidWorld tasks and reach the following success rates:

ModelCompleted TasksSuccess TasksSuccess Rate
gemini-3.8-flash116/1166656.90%
gpt-6.1-sol116/1165850.43%
claude-opus-5116/1164337.07%

These results contextualize Infinite-Actor's performance against recent proprietary models. Because these models are not open source, we cannot train them on our synthesized data and verify whether they would obtain the same gains. The experiments in the paper already provide sufficient evidence for the effectiveness of the proposed method: under the same backbone and evaluation protocol, Infinite-Actor improves over Qwen3-VL at 2B, 4B, and 8B on both AndroidWorld and MobileWorld.

3. Response to W1: Relative to Existing GUI World Models

Our key innovation is a unified state representation that bridges visual fidelity and action executability, which existing GUI world models do not jointly achieve. Text/HTML-based models support executable interactions but lack visual fidelity, while pixel-based simulators preserve visual appearance but lack structured representations of the visual changes induced by actions. In contrast, we represent GUI transitions as structured, coordinate-free deltas encoding trigger operations, added/removed elements, and subsequent affordances. This representation enables both learning from real interaction pairs and generation from text-only plans, bridging the gap between visually faithful simulation and executable GUI interactions. None of prior GUI world model has achieved such goal before:

Line of workState spaceConditioningTransition model
Qwen-AgentWorld, Code2Worldtext / HTMLaction + textLLM or HTML renderer
ViMo, NeuralOSpixelsactionvisual simulator trained from scratch
MobileDreamersketch abstractionactionsketch-based rollout model
Infinite-Dreamer (ours)pixelsstructured delta-textpretrained image editor fine-tuned on real pairs

we have revised our manuscript accordingly. (see PDF page 17, Appendix A.9; PDF page 2, Section 1; PDF page 3, Section 2)

4. Response to W2 and W4: Matched-budget Isolation and the Marginal Value of Imagination

We have conducted matched-budget comparisons at two levels. We first supplement the data-volume ablation to 27,360 samples, with a 9:1 real-to-synthetic ratio in each mixed group; data volume and optimization budget are matched exactly, and task coverage is matched at the task-suite level. Under this control, adding synthetic data improves AndroidWorld success to 44.83%, and Real+Aug+Gen reaches 48.28%, highlighting the complementary value of synthetic data:

ArmRealSyntheticPoolAW SR
Real only27,360027,36042.67
Real + Gen24,6242,73627,36044.83
Real + Aug24,6242,73627,36047.41
Real + Aug + Gen24,6242,73627,36048.28
Pure synthetic027,36027,36039.37

We further compare real and synthetic data using Qwen3.5-4B under matched training budgets. Synthetic data demonstrates strong scalability, surpassing real-data performance at a larger scale with lower cost:

Data budget0(baseline)1k2k3k6k
Real only37.5035.3443.5043.97-
Synthetic only37.5032.7639.2240.0945.69

Augmentation provides the most robust gain (+4.74), while multi-step imagination contributes +2.16 alone and +0.87 when combined with augmentation. A three-instance repeat gives +0.68 (95% CI [-0.32, 1.67], p=0.10) for imagination on top of augmentation, so this marginal gain is not resolved. Nevertheless, imagination uniquely expands state coverage beyond seed screenshots and enables longer successful episodes, highlighting its value beyond aggregate success rates.

ComparisonGen on realAug on realGen on AugFull mix on real
Delta+2.16+4.74+0.87+5.61

we have revised our manuscript accordingly. (see PDF page 8, Table 3 and Section 4.4; PDF page 46, Appendix I.1; PDF page 48, Appendix I.3)

5. Response to W3: Causes of the AC-Curated and Hard-task Regressions

We add a real-4B control under the same protocol, fine-tuning the base model on real data and comparing it with base-4B and Infinite-Actor-4B on AC-Curated-Hard. The regression has two sources: SFT and synthetic data contribute comparably to the Box Acc drop, while synthetic data contributes more to the Step SR drop. Thus, the regression stems partly from synthetic data and partly from fine-tuning itself:

ModelBox AccStep SR
base-4B77.0057.10
real-4B71.0352.31
Infinite-Actor-4B64.7044.20

We further analyze where the coordinate errors fall. The results show that the Hard median center distance grows from 0.18 to 0.38 annotated-box widths, 97% of incorrect boxes place their center more than 50 px outside the annotation, and Box Acc is 46.7% on the smallest third of targets against 75.4% on the largest third. The loss is therefore systematic: most errors are far outside the annotation and concentrate on compact targets. This explains why the offline metric drops while online task success improves at all three scales:

DifficultyMetricBaseInfinite-Actor
EasyMedian center distance0.060.13
HardMedian center distance0.180.38
EasyWrong boxes>50 px outside—88%
HardWrong boxes>50 px outside—97%
HardBox Acc, smallest / largest third67.5/87.046.7/75.4

Hard tasks require long, strictly ordered interactions. The clean long-horizon synthetic corpus is shallow: 84.2% of Hard tasks have a clean prefix of at least 5 steps, and 10.5% have a clean prefix of at least 10 steps (Appendix E.2). The mix therefore does not add enough long-horizon Hard supervision. The Hard difference is 10.5% vs. 13.2% (2.0 vs. 2.5 successful tasks out of 19); a task-level paired bootstrap gives a 95% CI of [-21.05, +15.79], so the difference is not resolved. The failure logs show budget exhaustion and false completion, consistent with insufficient long-horizon depth. These results show that the regression is a coordinate-precision trade-off rather than a general degradation, and they identify what to fix: coordinate-aware losses and more clean long-horizon data.

we have revised our manuscript accordingly. (see PDF page 8, Section 4.3; PDF pages 44-46, Appendix G.3; PDF pages 46-47, Appendix I.1; PDF pages 48-49, Appendix I.5)

6. Response to W5: Long-horizon Retention and Coverage of Difficult Workflows

We analyze data retention under the step-level filter. Results show that the 9.3% figure is a full-prefix diagnostic: 82.2% of generated steps are retained and 46.6% are clean, and at 10 or more steps, 12.8% of generated steps are retained and 1.0% are clean, so the all-pass rate understates retention by about 9x.

HorizonAt-riskRetainedCleanContaminatedPass rate
550.0%6.6%3.3%3.3%79.0%
828.8%3.9%0.9%2.9%79.4%
1018.9%2.5%0.3%2.2%79.5%
≥5—37.0%9.4%27.5%—
≥8—19.9%2.5%17.5%—
≥10—12.8%1.0%11.8%—
All—82.2%46.6%35.5%82.2%

We further report the coverage of difficult workflows. Results show that the retained corpus covers all 116 tasks, with 84.2% of Hard tasks containing at least one clean prefix of 5+ steps and 10.5% reaching 10+ steps:

ScopeTasks with clean ≥ 5 steps (%)Tasks with clean ≥ 10 steps (%)
All 11681.9%22.4%
Hard 1984.2%10.5%

These results show that the pipeline does produce usable long trajectories, and that the limitation is depth at the longest horizons rather than absence of difficult workflows.

we have revised our manuscript accordingly. (see PDF pages 36-37, Appendix E.2; PDF pages 46-47, Appendix I.1)

7. Response to W6: Visual Fidelity vs Executable Correctness

We add a paired action-recovery diagnostic (Appendix E.3) to assess action-state consistency. Backward recovery is nearly on par with real data, while forward recovery exhibits lower coordinate precision despite preserving action-type accuracy, indicating that the main gap lies in geometric localization rather than action semantics:

TestType accCoord hit@5%Box hitStep success
AC forward, real81.767.280.257.4
AC forward, generated72.729.734.833.9
AC backward, real93.771.888.662.2
AC backward, generated90.471.186.859.3

The forward localization gap is concentrated on clicks, while other action types retain stronger recovery performance:

ActionRealGenerated
click61.924.6
status69.158.1
wait50.041.1
input_text81.171.7

These findings establish synthetic data as a complement to real demonstrations for action semantics, appearance robustness, and state coverage, rather than a full replacement for coordinate supervision. Importantly, downstream agents trained with our synthetic data improve AndroidWorld and MobileWorld Pass@1 across all three scales (Table 1).

we have revised our manuscript accordingly. (see PDF page 6, Table 1; PDF pages 36-37, Appendix E.3; PDF page 48, Appendix I.4)

Thank you for the thorough rebuttal and additional analyses. I appreciate the authors' efforts in revising the manuscript and will update my score accordingly.

Reviewer_mjcoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Infinite-Dreamer, a framework for generating training data for GUI agents. It uses a VLM to extract structured descriptions of UI changes, called delta-text, from real screenshot transitions, then fine-tunes an image-editing model to generate subsequent screenshots conditioned on these descriptions. The framework produces both single-frame visual augmentations and multi-step imagined trajectories planned by an LLM. After VLM-based quality filtering, the synthetic data is used to train Infinite-Actor agents, improving task success on AndroidWorld and MobileWorld.

Critical strengths
  1. The paper addresses a practically important problem, as constructing high-quality training data for GUI agents is time-consuming and labor-intensive.
  2. The proposed approach is interesting and intuitive. This paper formulates GUI transitions as image-editing tasks and uses VLMs to describe action-induced UI changes as structured delta-text, enabling controllable synthesis of screenshot transitions.
Critical weaknesses
  1. The gap between synthetic and real GUI interaction data remains insufficiently characterized. Although Table 3 provides a useful downstream comparison, it is unclear whether data volume and task coverage are matched across settings, making it difficult to isolate differences in data quality. Controlled comparisons under matched data and training budgets, together with measurements of action-label accuracy and transition consistency, would better establish how effectively synthetic trajectories can substitute for or complement real demonstrations.

  2. The practical cost–benefit trade-off remains unclear. The full pipeline improves AndroidWorld success by 5.61 percentage points over real-data training, but requires additional world-model training, generation, and filtering. Reporting total costs and comparisons with cheaper augmentation methods would help establish whether these gains justify the overhead.

Questions for authors

See Weakness.

4 comments

Thanks for recognizing our contribution and novelty. We hope the following responses address your concern:

1. Response to W1: Gap of synthetic and real data

We conduct additional experiments under matched data volumes, task coverage, and training budgets. Specifically, we fine-tune Qwen3.5-4B on real and synthetic data at budgets of 1k, 2k, 3k, and 6k. While real data performs better at smaller budgets, 6k synthetic samples achieve a higher success rate at a substantially lower cost, demonstrating the scalability and cost-effectiveness of synthetic data:

Data budget0(baseline)1k2k3k6k
Real only37.5035.3443.5043.97-
Synthetic only37.5032.7639.2240.0945.69

We further assess synthetic trajectory quality through a paired action-recovery diagnostic (Appendix E.3). Reults show that backward recovery is nearly on par with real data, while forward recovery preserves action-type accuracy but exhibits lower coordinate precision, suggesting that the remaining gap lies in geometric localization rather than action semantics:

TestType accCoord hit@5%Box hitStep success
AC forward, real81.767.280.257.4
AC forward, generated72.729.734.833.9
AC backward, real93.771.888.662.2
AC backward, generated90.471.186.859.3

Finally, we compare real, synthetic, and mixed data training under matched data volumes (27,360 samples). Resutls show that adding synthetic data improves AndroidWorld success to 44.83%, with Real+Aug+Gen achieving 48.28%, highlighting the complementary value of synthetic data:

ArmRealSyntheticPoolAW SR
Real only27,360027,36042.67
Real + Gen24,6242,73627,36044.83
Real + Aug24,6242,73627,36047.41
Real + Aug + Gen24,6242,73627,36048.28
Pure synthetic027,36027,36039.37

we have revised our manuscript accordingly. (see PDF page 8, Section 4.4 and Table 3; PDF pages 36-37, Appendix E.3; PDF page 46, Appendix I.1; PDF page 48, Appendix I.3)

2. Response to W2: Cost–benefit

We conducted a cost analysis covering world-model fine-tuning, generation, filtering, and downstream training. Our pipeline costs only $0.032 per step, 9.6–11.9× less than OpenMobile and manual collection from AndroidControl and AITZ. This efficiency stems from offline, parallelizable generation, requiring only one diffusion pass and lightweight filtering per step, without emulator execution or weak-model interaction. World-model fine-tuning incurs a one-time cost of $432, which is independent of data volume.

OpenMobileCostCode2WorldCostOursCost
emulator$0.1316world model$0.0205Diffusion$0.0290
Monitoring$0.1629Action generation$0.0071Quality filtering$0.0013
Weak Model$0.0110Quality filtering$0.0029App Designer$0.0008
-————text-only delta$0.0010
Total Per Step$0.306/0.066hTotal Per Step$0.0306/0.057hTotal Per Step$0.032/0.019h
DatasetAndroidControl[1]AITZ[1]
Cost$0.377$0.345

End-to-end experiments reveal that although real data performs better at the same data volume, our approach achieves a higher success rate with substantially lower total cost at a larger scale:

MethodDataBudgetTotalCostAndroidWorld(SRPASS@1)
OpenMobile3k$940.03/202.75h43.97
Code2World2k$69.9/116.05h37.07
Ours6k$196.61/116.74h45.69

These results demonstrate the cost-effectiveness of our approach for scaling GUI agent training.

We have revised our manuscript accordingly. (see PDF pages 47-48, Appendix I.2)

[1] Rui Lv, et al. "M$^2$-Miner: Multi-Agent Enhanced MCTS for Mobile GUI Agent Data Mining" ICLR 2026

Thank you very much for your comment. We have supplemented relevant experiments to compare the gap between synthetic data and real data, and reported the required costs:

  1. Synthetic data is scalable, cost-effective, and complementary to real data Cost Report show that our pipeline costs only $0.032 per step, 9.6–11.9× less than OpenMobile and manual collection from AndroidControl and AITZ and supplementary experiments show that synthetic data can achieve a higher success rate than real data at substantially lower cost(45.69, $196.61 vs 43.97, $940.03), demonstrating its scalability and cost-effectiveness. Moreover, under the same data budget, adding synthetic data improves AndroidWorld success to 44.83%, and Real+Aug+Gen reaches 48.28%, highlighting the complementary value of synthetic data.

  2. Our core innovation is a unified state representation Our core innovation is a unified state representation that simultaneously accounts for visual realism and operational feasibility, something existing GUI world models cannot achieve. we represent GUI transitions as structured, coordinate-free deltas encoding trigger operations, added/removed elements, and subsequent affordances. This representation enables both learning from real interaction pairs and generation from text-only plans, bridging the gap between visually faithful simulation and executable GUI interactions. None of prior GUI world model has achieved such goal before.

We hope this response has addressed your questions. Please feel free to raise any further issues at any time; we are happy to continue the discussion.

Reviewer_mjcoOpenReview ↗

Official Comment by Reviewer

Thanks for your rebuttal. Overall, these additions address most of my original concerns. I have one remaining question on cost. The authors report a one-time world-model fine-tuning cost of $432, while the end-to-end cost for generating 6k samples is reported as $196.61. Could the authors clarify whether the reported end-to-end cost includes this one-time fine-tuning cost? If the authors can clarify this point, I would be happy to increase my score accordingly.

Thank you for the helpful question and for giving us the opportunity to clarify this point. We apologize for the lack of clarity in our previous response.

The reported end-to-end cost of $196.61 does not include the one-time world-model fine-tuning cost of $432. Notably, even if we add the training cost, our method remains more cost-effective and cheaper than OpenMobile, achieving a lower total cost ($628.61 vs. $940.03) while delivering a higher success rate (45.69% vs. 43.97%).

We will make this cost breakdown explicit in the revised manuscript. We sincerely thank the reviewer for recognizing our contributions and for the thoughtful feedback. We truly appreciate your support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.