acceptodds
Under review as a conference paper at ICLR 2027

GUI-ECO: An Efficient Edge-Cloud Orchestration Framework for GUI Agents via Dual Decision Processes

Abstract

The growing deployment of GUI agents on resource-constrained devices motivates collaboration between lightweight edge agents and more capable cloud agents. However, due to the multimodal and long-horizon nature of GUI tasks, existing collaboration frameworks struggle to determine when cloud assistance is needed and lack cloud-side acceptance mechanisms, resulting in low overall task success rates and unnecessary computational costs. To address these challenges, we propose **GUI-ECO**, a dual-decision **E**dge–**C**loud **O**rchestration framework, which comprises two modules: a trajectory-level edge handoff module and a competence-aware cloud acceptance module. We construct a dataset from extensive real-world GUI task executions and use the resulting supervision to train and refine both modules based on agents’ hidden representations. For edge-side handoff decisions with GUI-Owl-1.5, GUI-ECO achieves 85% Recall and 7% FPR on OSWorld. Overall, extensive experiments show that GUI-ECO achieves the highest cloud execution efficiency through its dual-decision process, reaching up to 2.6× that of prior methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
GUI, Agent, Edge-Cloud-Orchestration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GwmzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how small edge models and large cloud models cooperate on GUI agent tasks. The paper adds two checks: an edge module that tracks multi-step performance drops, and a cloud gate that rejects hopeless tasks early. Experiments show higher success rates while lower cloud step usage.

Critical strengths

The proposed method is intuitive.

Experiments show a performance gain.

Critical weaknesses
  1. Although the performance gain is valid, and the proposed method is intuitive, this work is generally an engineering work instead of a research work.

  2. The decision modules and LoRA weights are trained on the hidden states and trajectories of a specific GUI agent, so they cannot generalize to new agents or different failure patterns without retraining from scratch.

  3. Collecting training labels for each agent requires running 5 rollouts per intermediate state, which is very expensive to repeat whenever the base agent changes.

Questions for authors

Please address the weaknesses.

3 comments

W1: Research Contributions

We appreciate the reviewer’s recognition of the performance gains and the intuition behind GUI-ECO. We respectfully clarify that our contribution goes beyond integrating existing components: GUI-ECO investigates how task-completion competence can be learned from agents’ internal representations and used to guide edge--cloud collaboration. Specifically, our work makes the following contributions:

1. Formulating collaboration as two distinct decisions. We distinguish whether the edge agent needs assistance from whether the cloud agent is likely to complete the remaining task. GUI-ECO explicitly models both through edge-side handoff and cloud-side acceptance, rather than treating a request for assistance as sufficient justification for cloud execution. This formulation allows us to investigate when cloud intervention is beneficial and when further execution is unlikely to justify its cost.

2. Investigating temporal evidence for handoff decisions. We study whether the evolution of competence scores provides useful information beyond an individual state-wise estimate. This addresses a substantive decision-making problem: a temporary error or low score does not necessarily imply that the agent cannot recover and complete the task.

3. Learning agent-specific task-completion competence. We investigate whether an agent’s internal representations can predict its likelihood of completing the remaining task from the current GUI state. Repeated continuation rollouts provide outcome-grounded supervision for this prediction. Unlike directly predicting routing labels, our approach explicitly estimates the agent’s future task-completion probability from its current state, rather than task difficulty or individual action correctness.

4. Adapting representations for competence estimation. We further investigate whether competence information can be made more accessible through joint head–model training, rather than only extracted from frozen representations. LoRA adapts the representations consumed by the SCE head, BCE anchors predictions to empirical rollout success rates, and pairwise margin enhancement encourages appropriately ordered and separated scores across competence levels.

5. Evaluating these design choices through controlled experiments. Our evaluation across three agent pairs and four desktop/mobile benchmarks examines the task-success–cloud-execution trade-off under different model capabilities and environments. Complementary ablations investigate the roles of joint training, trajectory-level modeling, and cloud-side acceptance, while case studies illustrate how these mechanisms affect individual execution trajectories.

Taken together, these contributions concern problem formulation, competence representation learning, and sequential decision-making, rather than implementation alone.

W2: Agent-Specific Adaptation and Generalization Across Different Patterns

Although our decision modules and LoRA adapters are agent-specific, adapting GUI-ECO to a new agent requires only a modest model-training workload; the data-collection overhead is discussed separately in our response to W3. We collect training data from only two datasets and the additional components are lightweight and efficient to train, with their sizes reported in Table 4 of the original submission. For our largest evaluated agent, GUI-Owl-1.5-32B, LoRA adaptation takes approximately one hour on an NVIDIA A100 GPU. Furthermore, our OOD evaluations on MobileWorld and VisualWebArena show that the trained modules generalize across datasets and tasks without retraining, suggesting that they capture transferable competence and failure patterns rather than merely fitting the training trajectories.

W3: Rollout Data Collection Cost

We acknowledge that collecting five continuation rollouts per sampled state introduces an agent-specific data-collection cost. However, we consider this a manageable offline investment for operator-controlled cloud deployments. Such data collection is performed offline and can be automated and parallelized: for example, RoTS[1], developed by Alibaba Cloud Computing, hosts GUI environments on elastic cloud infrastructure and uses a rollout manager to coordinate asynchronous parallel sampling across distributed environment and agent servers.

Our data collection is restricted to sampled intermediate states from OSWorld and AndroidWorld, without requiring additional supervision for each deployment environment. Even for our largest evaluated agent, GUI-Owl-1.5-32B, collecting the complete rollout dataset required only approximately 100 NVIDIA A100 GPU-hours, representing a relatively modest and manageable offline cost for cloud service providers.

We sincerely hope that our responses can adequately address your concerns.

[1] Tianpeng Bu et al. Recovering Policy-Induced Errors: Benchmarking and Trajectory Synthesis for Robust GUI Agents. ICML, 2026.

Reviewer_GwmzOpenReview ↗

thanks for your response

i will maintain my original rating

Thank you for your feedback. We appreciate your time and consideration.

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

This paper proposes GUI-ECO for edge-cloud collaboration of GUI agents. On the edge side, it first estimates state-level task-completion competence from the model’s hidden states, and then uses a GRU to aggregate the sequence of competence scores to decide when to request cloud takeover; on the cloud side, a separate competence estimation head decides whether to accept the handoff. The competence estimator is trained with rollout success rates obtained through repeated continuation rollouts from intermediate states, after which the model and the estimation heads are jointly fine-tuned. However, the current version still has several issues that may affect the ICLR acceptance decision.

Critical strengths
  1. The problem setting is well aligned with the characteristics of GUI tasks. Transient errors in long-horizon trajectories do not necessarily imply failure of the entire task; moreover, deciding only on the edge side whether to hand off cannot determine whether the cloud side is capable of completing the remaining task. The two decision points are well motivated.
  2. The dual-decision design is simple, interpretable, and effective. Compared with performing only edge routing, adding a cloud acceptance gate is a natural but important improvement. The “w/o Cloud SCE Head” ablation in the experiments shows that removing this gate usually improves SR slightly but significantly increases cloud steps, indicating that this module indeed plays a role in cost control.
Critical weaknesses
  1. The method requires access to the hidden states of both the edge and cloud models, and also applies LoRA to the cloud model and trains an SCE head on it. In practice, mainstream cloud GUI agents are mostly closed source APIs (with no access to hidden states). The authors need to explicitly discuss this limitation.
  2. The paper essentially defines cost as cloud execution steps. This can capture part of the cloud inference computation, but it is insufficient to support conclusions about real edge-cloud efficiency. A more accurate phrasing at present should be “reduce cloud decoding/execution steps,” rather than directly equating this with full edge-cloud system efficiency.
  3. Cloud rejection directly triggering Early Stop makes the action spaces of the comparison not fully identical. In the paper's dual decision, if the edge side believes that a handoff is needed but the cloud SCE judges that it is also not reliable enough, the task is terminated immediately. This can itself be a reasonable strategy, but it raises an important comparison issue: GUIECO is given the action of "abandoning difficult tasks early to save cloud cost," whereas the policy space of most baselines is to continue on the edge or transfer to the cloud.
Questions for authors
  1. Hera is precisely aimed at step-level device-cloud coordination for longhorizon agents and explicitly optimizes task success and cloud usage. Although its task environments are not entirely consistent with the GUI benchmarks in this paper, it is highly relevant to the core problem of this paper. Currently, it is only cited generically in the Introduction, but it is not adequately discussed in Related Work, nor is there an experimental comparison or an explanation of why such a comparison cannot be conducted.
  2. The ablation on joint training is not sufficiently fine-grained. The w/o Joint Training setting simultaneously removes several factors: the adaptation of hidden representations via LoRA, the joint BCE, and the pairwise margin enhancement. Therefore, the current ablation only demonstrates that the full joint training pipeline is useful; it cannot establish whether the pairwise loss specifically emphasized in the paper is actually necessary.
  3. Table 3 shows that after removing joint training, NCSE drops from 2.84 to 0.74 (lower than the 1.00 of All-Cloud), indicating that the vast majority of the gains come from joint training rather than from the "dual decision" process itself that the paper highlights.
  4. The authors could report, for each benchmark, GUI-ECO’s handoff rate, cloud-side acceptance rate, rejection rate, distribution of handoff steps, and the average number of cloud execution steps required after acceptance. These metrics could better reflect the effectiveness of GUIECO.
  5. What is the criterion for selecting the thresholds δe\delta_e and δc\delta_c?
  6. The notation in Figure 2 is confusing and difficult to understand.

[1] Hera: Learning long-horizon coordination for devicecloud collaborative llm agents

4 comments

RESPONSE 1/3

W1-W3

W1: Cloud Model Accessibility and Deployment Limitations

We thank the reviewer for raising this concern. Our intended setting is operator-controlled edge--cloud deployment, where the system provider has access to the cloud model and can perform LoRA adaptation and train the corresponding SCE head. The cloud-side assessment can operate entirely within the provider’s infrastructure, without exposing hidden states to the edge client. We have explicitly clarified this deployment assumption and acknowledged the limited compatibility with black-box cloud services in the Limitations section (lines 1066–1069 of the revised manuscript).

W2: Cloud Computation Cost and Efficiency Measurement

We agree that cloud execution steps do not capture all cloud-side inference computation or the full cost of an edge--cloud system. End-to-end cost depends on model size, hardware configuration, inference implementation, and communication conditions, making comparisons sensitive to deployment-specific factors. We therefore use cloud execution steps as a hardware-independent proxy for cloud execution workload, providing a consistent and reproducible basis for comparing orchestration strategies within each fixed agent pair.

To complement this proxy, we additionally report average cloud-side computation in TFLOPs per task for GUI-Owl-1.5 (2B/32B) on OSWorld. GUI-ECO consumes 101.8 TFLOPs per task, compared with 1,659.3 TFLOPs for All-Cloud, corresponding to a 93.9% reduction.

MethodCloud Compute(TFLOPs/task)↓
All-Cloud1,659.3
All-Edge0.0
Random814.5
RouteLLM769.7
LightAgent1,082.0
MAI-UI618.5
Hera729.7
GUI-ECO101.8

W3: Justification for Early Termination

Early termination is an intentional capability of our dual-decision process. The edge module determines whether cloud assistance is needed, while the cloud SCE head assesses whether the cloud agent is likely to complete the remaining task. When the estimated cloud competence is insufficient, GUI-ECO can terminate execution early to avoid unnecessary cloud computation. This decision stems from the learned competence assessment of the cloud SCE head, rather than an arbitrary mechanism for discarding difficult tasks, representing a key advantage of our dual-decision design. Importantly, early-terminated tasks are counted as failures in the overall success rate, and the computational overhead of cloud-side SCE assessment is also accounted for in our cost analysis.

To address the concern that early termination may reduce cloud computation by abandoning potentially solvable tasks, we perform additional cloud continuation rollouts from rejected handoff states and report the proportion of rejected requests that succeed under continued cloud execution. The results are summarized below.

Agent PairBenchmarkRejectedFalsely RejectedFalse Rejection Ratio ↓
GUI-Owl-1.5 (2B/32B)OSWorld2713.70%
VisualWebArena47510.64%
AndroidWorld1417.14%
MobileWorld3837.89%
MAI-UI (2B/8B)OSWorld3625.56%
VisualWebArena4237.14%
AndroidWorld1915.26%
MobileWorld35411.43%
UI-Mate (9B/27B)OSWorld2414.17%
VisualWebArena47612.77%
AndroidWorld2514.00%
MobileWorld50510.00%
TotalAll Benchmarks404338.17%

Across all three agent pairs and four benchmarks, only 33 out of 404 rejected requests (8.17%) can be successfully completed by the cloud agent, while the remaining 91.83% fail despite further cloud execution. These results indicate that our cloud-side acceptance mechanism effectively identifies requests with low task-completion potential, rather than simply abandoning difficult tasks to reduce costs.

RESPONSE 2/3

Q1-Q3

Q1: Comparison with Hera

We thank the reviewer for highlighting Hera and agree that it is highly relevant to our work. Hera employs an external coordinator to select edge or cloud execution at every interaction step. Our original evaluation primarily focused on task-level allocation and one-way handoff protocols. In GUI-ECO, the edge agent continuously monitors execution and determines when to request a handoff; once accepted, the cloud agent takes over the remaining execution. This difference in orchestration protocols motivated our initial experimental scope. We have also added a discussion of Hera to the Related Work section (lines 133–136 of the revised manuscript).

Despite these differences in edge--cloud orchestration protocols, we adapt Hera to the GUI-agent setting to provide a more comprehensive evaluation. We compare it with GUI-ECO using the same agent pairs and training/test task splits. The results are summarized in the table below. Methodologically, Hera directly learns step-level model-selection decisions, with training that incorporates trajectory returns and future cloud usage. In contrast, GUI-ECO learns explicit, outcome-grounded competence estimates from the agents’ internal representations, providing a more direct and interpretable basis for handoff decisions.

Agent PairBenchmarkMethodSR ↑ (%)Cloud Steps ↓NCSE ↑
GUI-Owl-1.5 (2B/32B)OSWorldGUI-ECO48.531.598.26
Hera48.534.383.00
VisualWebArenaGUI-ECO41.674.991.67
Hera39.177.180.96
AndroidWorldGUI-ECO66.670.6511.39
Hera64.812.372.08
MobileWorldGUI-ECO32.385.121.66
Hera31.438.290.77

Q2: Fine-Grained Ablation of Joint Training Components

We thank the reviewer for this suggestion and agree that the original ablation does not isolate the contribution of each training component. We therefore conduct finer-grained ablations by separately removing LoRA adaptation, the BCE objective during joint training, and pairwise margin enhancement. As shown in the table below, removing any of these components reduces NCSE, with the largest degradation observed when LoRA is removed. These components serve complementary roles: LoRA adapts the hidden representations to make task-completion competence more accessible to the SCE head; pairwise margin enhancement encourages appropriately ordered and separated scores for states with different competence levels; and BCE anchors the competence predictions to empirical rollout success rates throughout adaptation.

VariantOSWorld SR ↑ (%)Cost ↓VisualWebArena SR ↑ (%)Cost ↓AndroidWorld SR ↑ (%)Cost ↓MobileWorld SR ↑ (%)Cost ↓NCSE ↑
GUI-ECO48.531.5941.674.9966.670.6532.385.122.84
w/o LoRA44.123.4937.506.6662.961.5429.527.340.95(-1.89)
w/o Joint BCE45.593.5038.336.8064.811.6730.487.701.06(-1.78)
w/o Pairwise Margin45.593.5438.336.7062.961.5030.487.601.18(-1.66)
VariantOSWorld SR ↑ (%)Cost ↓VisualWebArena SR ↑ (%)Cost ↓AndroidWorld SR ↑ (%)Cost ↓MobileWorld SR ↑ (%)Cost ↓NCSE ↑
GUI-ECO26.473.1443.334.9155.562.0920.007.321.70
w/o LoRA23.534.9739.176.9353.703.9617.1410.260.79(-0.91)
w/o Joint BCE25.004.5740.836.5353.703.5619.059.541.02(-0.68)
w/o Pairwise Margin25.004.8240.006.8053.703.8118.109.970.90(-0.80)

Q3: Relationship Between Joint Training and Dual-Decision Design

We agree that joint training makes a substantial contribution to GUI-ECO. However, joint training and the dual-decision process are complementary rather than independent components. Joint training improves the competence estimates that underpin both decisions: the edge EDM learns from sequences of edge competence scores, while the cloud acceptance gate directly uses cloud competence estimates.

As shown in Figure 4 of paper, joint training improves competence-estimation recall and reduces FPR, providing more reliable signals for downstream handoff and acceptance decisions. Consequently, removing joint training degrades the inputs to both decision modules. The substantial NCSE decrease therefore highlights the importance of reliable competence estimation to the dual-decision process; it does not, by itself, establish that joint training accounts for most of the gains independently of that process.

RESPONSE 3/3

Q4-Q6

Q4: Additional GUI-ECO Statistics

The two tables below provide additional statistics across all three agent pairs and four benchmarks. The first reports handoff rates, cloud-side acceptance and rejection rates, and average cloud execution steps over all tasks and accepted tasks. Acceptance and rejection rates are calculated over handoff requests. The second presents the distribution of the steps at which handoffs are first triggered, including both accepted and rejected requests. Together, these statistics clarify when cloud assistance is requested, how selectively requests are accepted, and the subsequent cloud execution workload.

Agent pairBenchmarkHandoff rateAcceptance rateRejection rateCloud Steps / All TasksCloud Steps / Accepted Task
GUI-Owl-1.5 (2B/32B)OSWorld52.94%(36/68)25.00%(9/36)75.00%(27/36)1.5912.01
GUI-Owl-1.5 (2B/32B)VisualWebArena65.83%(79/120)40.51%(32/79)59.49%(47/79)4.9918.71
GUI-Owl-1.5 (2B/32B)AndroidWorld37.04%(20/54)30.00%(6/20)70.00%(14/20)0.655.85
GUI-Owl-1.5 (2B/32B)MobileWorld64.76%(68/105)44.12%(30/68)55.88%(38/68)5.1217.92
MAI-UI (2B/8B)OSWorld69.12%(47/68)23.40%(11/47)76.60%(36/47)3.1419.41
MAI-UI (2B/8B)VisualWebArena64.17%(77/120)45.45%(35/77)54.55%(42/77)4.9116.83
MAI-UI (2B/8B)AndroidWorld48.15%(26/54)26.92%(7/26)73.08%(19/26)2.0916.12
MAI-UI (2B/8B)MobileWorld73.33%(77/105)54.55%(42/77)45.45%(35/77)7.3218.30
UI-Mate (9B/27B)OSWorld54.41%(37/68)35.14%(13/37)64.86%(24/37)4.3022.49
UI-Mate (9B/27B)VisualWebArena61.67%(74/120)36.49%(27/74)63.51%(47/74)3.9917.73
UI-Mate (9B/27B)AndroidWorld68.52%(37/54)32.43%(12/37)67.57%(25/37)3.2714.72
UI-Mate (9B/27B)MobileWorld70.48%(74/105)32.43%(24/74)67.57%(50/74)2.5811.29
Agent pairBenchmarkHandoff CountSteps 1–5Steps 6–10Steps 11–20Steps >20
GUI-Owl-1.5 (2B/32B)OSWorld366(16.67%)13(36.11%)12(33.33%)5(13.89%)
GUI-Owl-1.5 (2B/32B)VisualWebArena7916(20.25%)30(37.97%)24(30.38%)9(11.39%)
GUI-Owl-1.5 (2B/32B)AndroidWorld207(35.00%)8(40.00%)4(20.00%)1(5.00%)
GUI-Owl-1.5 (2B/32B)MobileWorld688(11.76%)21(30.88%)28(41.18%)11(16.18%)
MAI-UI (2B/8B)OSWorld479(19.15%)17(36.17%)15(31.91%)6(12.77%)
MAI-UI (2B/8B)VisualWebArena7713(16.88%)28(36.36%)26(33.77%)10(12.99%)
MAI-UI (2B/8B)AndroidWorld268(30.77%)11(42.31%)6(23.08%)1(3.85%)
MAI-UI (2B/8B)MobileWorld779(11.69%)23(29.87%)31(40.26%)14(18.18%)
UI-Mate (9B/27B)OSWorld375(13.51%)13(35.14%)14(37.84%)5(13.51%)
UI-Mate (9B/27B)VisualWebArena7411(14.86%)26(35.14%)27(36.49%)10(13.51%)
UI-Mate (9B/27B)AndroidWorld379(24.32%)14(37.84%)11(29.73%)3(8.11%)
UI-Mate (9B/27B)MobileWorld747(9.46%)21(28.38%)31(41.89%)15(20.27%)

Q5: Threshold Selection Criteria

We thank the reviewer for requesting this clarification. For each agent pair, both thresholds are selected exclusively on the validation set by maximizing Recall subject to an FPR of at most 10%. Specifically, the edge handoff threshold δe\delta_e is calibrated using task-level handoff Recall and FPR, while the cloud acceptance threshold δc\delta_c is calibrated using state-level acceptance Recall and FPR under the same selection criterion. Once selected, both thresholds remain fixed during testing. We have added this clarification to lines 318–320 of the revised manuscript.

Q6: Clarification of Figure 2 Notation

We thank the reviewer for pointing this out. We have added annotations next to potentially ambiguous symbols in Figure 2—for example, gg denotes the GUI screen and cc denotes the prompt context. We hope these annotations make the notation clearer and the overall workflow easier to follow.

Dear Reviewer qZtu,

Thank you again for your thoughtful feedback. As the discussion period draws to a close, we would appreciate your feedback on our response. If it has satisfactorily addressed your concerns, we would be grateful if you would consider revising your score upward. Should any questions remain, we would be happy to provide further clarification.

Thank you for your time and consideration.

Best regards,

The authors of Paper GUI-ECO: An Efficient Edge-Cloud Orchestration Framework for GUI Agents via Dual Decision Processes

Reviewer_U1b7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents GUI-ECO, a framework where edge and cloud GUI agents work together to raise task success rates while reducing cloud compute costs. It relies on a dual-decision Edge–Cloud Orchestration framework, which comprises two modules: a trajectory-level edge handoff module and a competence-aware cloud acceptance module. To support this, the authors build a dataset from repeated rollouts of task executions and train State-Level Competence Estimation (SCE) heads by optimizing the heads and the underlying model jointly. Tested on three pairs of edge and cloud agents across four GUI benchmarks, the approach uses cloud resources more efficiently than existing collaboration strategies.

Critical strengths
  1. The paper addresses a practical problem: balancing task success and cloud cost when deploying GUI agents across edge and cloud.
  2. The case studies are descriptive and clearly show how the framework decides when to hand off and when the cloud accepts, making the mechanism easy to follow.
Critical weaknesses
  1. GUI-ECO relies on fixed thresholds for edge handoff and cloud acceptance, which may not generalize well across different agents, tasks, and environments. The paper lacks sufficient analysis of threshold sensitivity and robustness, raising concerns about the generalizability of the proposed decision mechanism.
  2. Since the MAI-UI and LightAgent baselines are implemented using simplified or prompt-only monitoring mechanisms, can the authors provide comparisons with stronger learned routers under matched supervision and computation budgets?
  3. Figure 3 shows that GUI-ECO and MAI-UI achieve identical FPR (7.69%) with UI-Mate on AndroidWorld, contradicting the claim of consistently lower FPR. Although GUI-ECO achieves higher Recall, the authors should clarify this observation and provide further analysis of the robustness of the reported improvements.
Questions for authors
  1. GUI-ECO relies on fixed thresholds for edge handoff and cloud acceptance, which may not generalize well across different agents, tasks, and environments. The paper lacks sufficient analysis of threshold sensitivity and robustness, raising concerns about the generalizability of the proposed decision mechanism.
  2. Since the MAI-UI and LightAgent baselines are implemented using simplified or prompt-only monitoring mechanisms, can the authors provide comparisons with stronger learned routers under matched supervision and computation budgets?
  3. Figure 3 shows that GUI-ECO and MAI-UI achieve identical FPR (7.69%) with UI-Mate on AndroidWorld, contradicting the claim of consistently lower FPR. Although GUI-ECO achieves higher Recall, the authors should clarify this observation and provide further analysis of the robustness of the reported improvements.
2 comments

W1: Threshold Calibration and Robustness

We thank the reviewer for raising this concern. We clarify that the edge handoff threshold τe\tau_e and cloud acceptance threshold τc\tau_c are independently calibrated on the validation set for each agent pair, rather than fixed universally across agents or decision modules. Our validation data covers diverse desktop and mobile GUI tasks, and the calibrated thresholds transfer to two OOD benchmarks without recalibration. Such threshold calibration is also adopted in learned routing methods such as RouteLLM [1], which uses representative samples for calibration.

We assess threshold robustness on GUI-Owl-1.5 and MAI-UI by adding uniform noise in [-0.1,0.1] to the thresholds, corresponding to at most 10% of the sigmoid-normalized competence score range. The results are summarized below. GUI-ECO maintains comparable overall performance under the tested perturbations, with no substantial degradation in effectiveness. For example, the GUI-Owl-1.5 32B cloud-side decision module changes its decisions in only 2 of 203 evaluated cases (0.99%) relative to the unperturbed setting. These findings support the stability of our decision mechanism under small threshold variations and are consistent with joint training objective, which encourages greater separation between states with different competence levels.

Agent pairBenchmarkPerturbationsSR ↑(%)Cloud Steps ↓NCSE↑
GUI-Owl-1.5(2B/32B)VisualWebArenaW/O41.674.991.67
VisualWebArenaW/40.835.301.48 (-0.19)
MobileWorldW/O32.385.121.66
MobileWorldW/33.335.851.82 (+0.16)
MAI-UI(2B/8B)VisualWebArenaW/O43.334.911.27
VisualWebArenaW/41.675.031.12 (-0.15)
MobileWorldW/O20.007.321.29
MobileWorldW/20.007.801.21 (-0.08)

W2: Comparison with Stronger Learned Routers

We thank the reviewer for this suggestion. We would like to gently remind the reviewer that, in addition to the prompt-based implementations of MAI-UI and LightAgent, our evaluation includes RouteLLM [1], a learned routing approach that predicts the relative advantage of edge and cloud models to guide task allocation.

Furthermore, we extend our evaluation to Hera [2], a step-level routing framework that introduces an additional Qwen2.5-0.5B model trained through imitation learning and cost-aware reinforcement learning to select edge or cloud execution at each step. We adapt Hera to the GUI-agent setting and evaluate it against GUI-ECO using the same agent pairs and training/test task splits. The results are summarized in the table below. Methodologically, Hera directly learns step-level model-selection decisions, with training that incorporates trajectory returns and future cloud usage. In contrast, GUI-ECO learns explicit, outcome-grounded competence estimates from the agents’ internal representations, models their temporal evolution for edge-side handoff, and separately evaluates cloud-side acceptance, providing a more direct and interpretable basis for handoff decisions.

Agent PairBenchmarkMethodSR ↑ (%)Cloud Steps ↓NCSE ↑
GUI-Owl-1.5 (2B/32B)OSWorldGUI-ECO48.531.598.26
Hera48.534.383.00
VisualWebArenaGUI-ECO41.674.991.67
Hera39.177.180.96
AndroidWorldGUI-ECO66.670.6511.39
Hera64.812.372.08
MobileWorldGUI-ECO32.385.121.66
Hera31.438.290.77
MAI-UI (2B/8B)OSWorldGUI-ECO26.473.143.20
Hera22.065.760.69
VisualWebArenaGUI-ECO43.334.911.27
Hera42.507.260.81
AndroidWorldGUI-ECO55.562.096.60
Hera57.414.632.33
MobileWorldGUI-ECO20.007.321.29
Hera19.0510.180.77

W3: FPR Discrepancy and Robustness Analysis

We thank the reviewer for pointing out this imprecise statement. For UI-Mate on AndroidWorld, only 13 out of 54 tasks are successfully completed by the edge agent, resulting in a coarse FPR resolution of 7.69%. Both GUI-ECO and MAI-UI produce only one false positive, leading to identical FPR despite GUI-ECO achieving higher Recall. To further assess robustness, we expand the evaluation to 108 tasks, of which 29 do not require cloud assistance. GUI-ECO produces only 3 false positives, compared with 7 for MAI-UI, reducing FPR from 24.14% to 10.34%. We also revise our original claim to "lower or comparable FPR" for greater precision (Line 393).

[1] Isaac Ong et al. RouteLLM: Learning to Route LLMs from Preference Data. ICLR, 2025. [2] Yuxin Zhang et al. Hera: Learning Long-Horizon Coordination for Device-Cloud Collaborative LLM Agents. arXiv preprint arXiv:2605.24598, 2026.

Dear Reviewer U1b7,

Thank you again for your thoughtful feedback. As the discussion period draws to a close, we would appreciate your feedback on our response. If it has satisfactorily addressed your concerns, we would be grateful if you would consider revising your score upward. Should any questions remain, we would be happy to provide further clarification.

Thank you for your time and consideration.

Best regards,

The authors of Paper GUI-ECO: An Efficient Edge-Cloud Orchestration Framework for GUI Agents via Dual Decision Processes

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.