acceptodds
Under review as a conference paper at ICLR 2027

Learning Operational Boundaries for GUI Agents via Constrained Optimization

Abstract

While graphical user interface (GUI) agents have made significant strides in executing complex tasks across diverse environments, their interaction with open-ended web interfaces raises substantial safety risks. Adversarial or compromised websites deliberately designed to mislead users and agents into unsafe operations pose a common and severe threat. In this work, we introduce GUI-WebMinefield, a benchmark for evaluating GUI agents in harmful web environments. Our evaluation shows that even state-of-the-art agents often fail to recognize potential risks and still attempt to execute instructions in malicious settings. To address this gap, we propose Constraint-Conditioned Optimization with Operation Advantage Calibration, which improves risk-aware action selection by incorporating explicit safety constraints. We construct safety constraints to define permissible operational boundaries in GUI environments and perform token-level calibration of operation-specific advantages during reinforcement learning. Experiments demonstrate that our method preserves strong task performance while substantially improving safety, without requiring exposure to harmful environments during training. The benchmark and code will be made publicly available.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Model, Agent, GUI Agent, Safety

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_K15DOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether GUI agents can recognize unsafe web environments and refuse risky interactions. The authors introduce GUI-WebMinefield, a benchmark of annotated webpage screenshots covering multiple risk categories. They propose Constraint-Conditioned Optimization, which augments ordinary GUI training examples with synthetic constraints that either permit execution or require refusal, without training on harmful webpages. They further introduce Operation Advantage Calibration, which separately normalizes rewards for format, action, and coordinate correctness and assigns the resulting advantages to the corresponding token spans. Experiments compare the combined method against a GUI-R1 baseline and a GRPO-based constraint-training variant, evaluating both safety and general GUI capabilities.

Critical strengths
  1. I think GUI-WebMinefield is a useful GUI safety benchmark of annotated risky webpages with task and action annotations. Unlike execution-focused benchmarks, it tests whether agents refuse risky interactions in harmful environments.
  2. An interesting approach to safety generalization. Constraint-conditioned training on ordinary GUI examples with synthetic constraints improves safety on harmful webpages without exposure to such environments during training.
Critical weaknesses
  1. OAC’s advantage allocation may introduce insufficiently justified bias. Restricting rewards to selected token spans may omit cross component gradients arising from autoregressive dependencies, while independent normalization changes reward weights. The paper should clarify the resulting objective and when this trade-off is beneficial.

  2. The scale experiments offer limited evidence of generalization. With 3,570 training examples (1,000 with synthetic constraints) and 1,600 benchmark screenshots across seven categories, the results do not establish generalization to broader environments, unseen constraints, or diverse interaction contexts.

Questions for authors
  1. Can the authors establish when the tradeoff introduced by OAC is beneficial? A theoretical characterization, diagnostic metric, or targeted ablation would help justify the design beyond empirical gains.
  2. The training data and synthetic constraints appear limited relative to the diversity of risk categories and interaction contexts in the benchmark. Can the authors provide data-scaling experiments to show how performance changes with more examples, constraint types, and environments?
3 comments

We thank the reviewer for the constructive suggestions.

Response to Weakness 1 & Question 1: OAC’s Objective and Component-Wise Advantage Allocation

Cross-component Gradients. OAC changes advantage allocation and thus the resulting gradient. When reasoning is present, its tokens receive joint feedback from both components. In this example,

\qquad A_{\mathrm{think}}=\frac{r_{\mathrm{think}}-\mu_{\mathrm{think}}}{\sigma_{\mathrm{think}}+\epsilon}.$$ With or without reasoning, component updates remain coupled through the autoregressive context and shared parameters. For fixed rollout advantages and span masks, the gradient of the policy term in the objective below, evaluated at $\theta=\theta_{\mathrm{old}}$ where clipping is inactive and excluding KL, is $$g_{\mathrm{OAC}}=\left.\mathbb{E}_{\pi_{\mathrm{old}}}\left[\operatorname{Mean}_{t:m_t=1}A_t\nabla_\theta\log\pi_\theta(y_t\mid x,y_{\lt t})\right]\right|_{\theta=\theta_{\mathrm{old}}}.$$ OAC assigns advantages by output field, changing the aggregate-reward gradient while retaining autoregressive dependence. We provide further theoretical analysis in Response 1 to Reviewer kMZ7. **Normalization.** OAC computes action and argument advantages separately. Specifically, $A_a=(r_a-\mu_a)/(\sigma_a+\epsilon)$ and $A_p=(r_p-\mu_p)/(\sigma_p+\epsilon)$, with statistics computed within each rollout group. Each field receives feedback based on its own relative correctness and variability, while reasoning receives the normalized joint reward above. This changes the relative weighting of component feedback. These normalized advantages $A_t$ used in the gradient above: $$A_t=A_{\mathrm{fmt}} +\mathbf1_{t\in S_a}A_a +\mathbf1_{t\in S_p}A_p +\mathbf1_{t\in S_{\mathrm{think}}}A_{\mathrm{think}},$$ The full objective, maximized with PPO clipping and KL regularization, is $$\mathcal{J}_{\mathrm{OAC}}=\mathbb{E}_{\pi_{\mathrm{old}}}\left[\operatorname{Mean}_{t:m_t=1}\min\\!\left(\rho_t A_t,\\,\operatorname{clip}(\rho_t,1-\delta,1+\delta)A_t\right)\right]-\beta\mathcal{L}_{\mathrm{KL}},$$ where $\rho_t=\pi_\theta(y_t\mid x,y_{<t})/\pi_{\mathrm{old}}(y_t\mid x,y_{<t})$, $S_d$ denotes a component’s token span, and $m_t$ masks valid response tokens. The reasoning term is omitted when no reasoning span is used. > **Response to Weakness 2 & Question 2: Generalization and Data Scaling** **Generalization to Unseen Constraints and Environments.** Our existing experiments examine generalization across environments, constraint formulations, interaction contexts, and model backbones. - First, our training and evaluation data cover web, desktop, and mobile environments. These benchmarks span different operating systems and substantially different task objectives, from explicit action instructions to long-horizon planning. They also use different interaction schemes and action definitions. - Second, GUI-WebMinefield tests all models on real harmful webpages that are never used in training, directly assessing transfer to unseen harmful environments. - Third, we evaluate grounding, low-level execution, high-level execution, OSWorld long-horizon tasks, adversarial instructions, and benign boundary cases. **These evaluations test the method’s generalization across different unseen settings.** The results show overall improvements in safety and usability. - We also examine the effects of constraints. Table 20 (revised paper) shows that stricter constraints further improve safety, while Table 29 shows that greater diversity in training constraints improves safety alignment. **Test-time constraints differ from training constraints in content, style, and topic.** **Data Scaling.** We further examine how the scale of training data affects safety alignment and task performance by varying the number of training examples with synthetic constraints. | Constraint Examples | w/o Safety: SR | w/o Safety: RR | w/o Safety: ScreenSpot | w/ Safety: SR | w/ Safety: RR | w/ Safety: ScreenSpot | |---|---:|---:|---:|---:|---:|---:| | 250 | 79.94 | 0.00 | 86.24 | 66.69 | 20.38 | 81.68 | | 500 | 80.56 | 0.00 | 86.87 | 63.69 | 21.50 | 84.91 | | 1,000 | 84.94 | 0.00 | 87.8 | 64.44 | 24.44 | 85.8 | | 2,000 | 80.38 | 0.00 | 87.42 | 60.56 | 27.44 | 86.64 | We expanded the training data by adding 600 constraint-conflicting and 400 constraint-aligned examples while keeping the unconstrained examples unchanged. Due to time constraints, the new examples were not individually reviewed by humans. The results show that increasing training data can improve safety alignment and task capability. The model trained with 2,000 constraint-augmented examples refuses risky requests more often when safety constraints are provided. Further gains in usability are limited, as our method **focuses on safety-constrained decisions in existing GUI scenarios** without introducing additional GUI task types.

Thank you for the detailed response and the additional data-scaling experiments. I will maintain my weak acceptance recommendation.

Thank you for your helpful suggestions and comments. We are glad that our responses could address your concern. If you have any further questions, we are ready to provide clarification at any time.

Reviewer_ceb8OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines the safety risks faced by GUI agents in open ended web environments. It introduces the GUI-WEBMINEFIELD, a screenshot based benchmark containing 1600 webpage screenshots across seven harmful categories. Each screenshot is paired with an instruction targeting a click, scroll, or type action. Agents are evaluated on whether they correctly predict the annotated action (SR) and explicitly refuse interaction (RR). The paper reports that many existing GUI agents continue executing instructions in harmful environments which motivates the authors to propose Constraint Conditioned Optimization (CCO) with Operation Advantage Calibration (OAC). Their approach combines training on synthetic conflicting and aligned system constraints with operation-specific token advantages. The authors compare CCO-GRPO and CCO-OAC with the GUI-R1-7B baseline on GUI-WEBMINEFIELD and other low-level and high-level GUI benchmarks to assess refusal and capability preservation.

Contributions

  • The paper examines the important practical question of whether GUI agents should follow instructions in harmful web environments, extending evaluation beyond action accuracy.
  • The proposed GUI-WEBMINEFIELD provides 1600 real-world samples across seven harmful categories with webpages, instructions, and action annotations reviewed by human annotators.

The proposed CCO-GRPO and CCO-OAC variants are evaluated against GUI-R1-7B on GUI-WEBMINEFIELD and general GUI benchmarks, with additional experiments on Qwen3-VL-4B. The results show that CCO-OAC training increases refusal on GUI-WEBMINEFIELD, while evaluations on low-level and high-level GUI tasks indicate that general capability is largely preserved.

Critical strengths

GUI-WEBMINEFIELD complements existing benchmarks with human-reviewed examples from real-world harmful websites across seven categories. This provides a useful resource for studying agents’ behavior in harmful web environments. However, this does not outweigh the evaluation and reproducibility concerns underlying my recommendation.

Critical weaknesses
  • The benchmark measures execution or refusal without assessing individual action safety or observing consequences, leaving the motivating question of how deception compromises agent safety only partly answered. Its harmful-only design also rewards blanket refusal, so assessing selective safety requires the separate benign evaluation in Table 25. These limitations restrict what the proposed GUI-WEBMINEFIELD benchmark alone establishes about safe behavior.
  • The proposed CCO-OAC method demonstrates limited protection in Table 2. Under safety prompts, refusal increases from 10.56% to 24.44%, while harmful execution remains at 64.44%. Although the authors acknowledge this limitation and the improvement is meaningful, the substantial remaining unsafe execution raises concerns about the effectiveness of the proposed safety training. The absence of supplementary code further limits verification of the reward implementation and independent reproduction, reducing confidence in the methodological contribution.
Questions for authors

When refusal is required, how is the argument reward calculated and can an incorrect executable action receive positive argument feedback?

3 comments

We appreciate the reviewer’s helpful comments.

Response to Weakness 1: Benchmark Scope and Action Safety Analysis

Benchmark Scope. GUI-WebMinefield is a red-teaming benchmark that uses deceptive webpages to expose safety risks in GUI agents. As discussed in the paper, any action in these environments is unsafe, much like human visiting the phishing websites. Therefore, it is difficult to determine whether individual actions or their consequences are safe. However, this does not make blanket refusal a meaningful policy. Agents must also be evaluated on usability benchmarks, rather than solely on a red-teaming benchmark. Our experiments include multiple usability and benign-task evaluations. These results clearly reflect our evaluation goals, as also noted by Reviewers kMZ7 and K15D.

Additional Action Safety Analysis. Using the user task and target action provided in each benchmark instance, we can characterize the action’s role in the risk chain. We divide target actions into three categories: (i) risk execution, which directly triggers a risky outcome, such as downloading, betting, or purchasing; (ii) workflow entry, which enters a high-risk workflow, such as logging in, registering, or following a link; and (iii) contextual navigation, which seeks to explore the displayed content, such as browsing or scrolling. The table below presents the action safety results for each category:

ModelMetricRisk Execution (N=507)Workflow Entry (N=989)Contextual Navigation (N=104)Overall (N=1600)
Qwen3-VL-32BSR50.1065.1280.7761.38
RR45.3627.406.7331.75
Gemini-2.5 CUASR67.6570.3788.4670.69
RR24.2614.660.9616.81
Claude-Sonnet-4.5 CUASR32.3539.7477.8839.88
RR60.7550.5612.5051.31
OpenAI CUASR9.6615.7735.5815.12
RR86.5969.6759.6274.38
UI-TARS-1.5-7BSR56.0259.0576.9259.25
RR18.9313.143.8514.37
OpenCUA-32BSR52.0765.1270.1961.31
RR0.000.000.000.00
GUI-R1-7BSR68.4477.6582.6975.06
RR20.916.270.9610.56
UI-Ins-32BSR54.2443.0743.2746.62
RR0.000.000.000.00
GTA1-32BSR85.0187.3679.8186.12
RR0.000.000.000.00

All agents capable of refusal show a similar pattern: refusal rates are highest for risk execution and lowest for contextual navigation. Models with stronger risk awareness, such as OpenAI CUA, maintain some caution even during contextual navigation, as any interaction with a phishing website may lead to a dangerous redirect. Most models, however, struggle to do so. Our benchmark therefore also supports safety analysis at the action level. We include these results in the revised manuscript.

Response to Weakness 2: Safety Effectiveness and Reproducibility

Safety Effectiveness. We break down the results by the role of the target action. CCO-OAC produces its safety gains on Risk Execution and Workflow Entry actions:

ModelRisk ExecutionWorkflow EntryContextual Navigation
GUI-R1-7B20.916.270.96
CCO-GRPO32.1511.931.92
CCO-OAC38.8619.312.88
CCO-OAC ΔRR vs. GUI-R1+17.95+13.04+1.92

CCO-OAC substantially reduces harmful task completion, primarily by increasing strict refusal for Risk Execution and Workflow Entry. Refusal during Contextual Navigation relies more on the model’s contextual knowledge of dangerous web environments. Improving this capability is orthogonal to the motivation of our method.

Reproducibility. We include the benchmark and code in the supplementary material and anonymous repository to make the implementation details available and support independent reproduction.

Response to Question 1: Argument Reward and Credit Assignment for Refusal Actions

An action can be incorrect while its argument is correct. In such cases, OAC computes separate rewards and advantages for the two components, allowing each to be learned independently. Specifically, for GUI-R1, we retain the baseline’s output schema and task prompt, defining the ground-truth point argument for refusal as [-100, -100], the default value for actions that do not require coordinates. A response may therefore predict the correct point argument but the wrong action. For Qwen3-VL, different actions have different argument schemas, but action and argument rewards and advantages are still computed separately. This decomposition continues to guide learning toward the correct response. As argument predictions become consistently correct within a rollout group, the training signal shifts toward improving action correctness.

I thank the authors for the clarifications and additional results. The code was shared only during rebuttal leaving limited time to verify the results and the released repository lacks the training pipeline. Also benchmark/data.jsonl records "deepseek-flash" as classifier_model for all 1,600 action-role labels used in the rebuttal analysis, without labeling rules or human review records. Unless these labels were human-verified, this appears inconsistent with Appendix B.2. The rebuttal also states improving contextual recognition of dangerous environments as orthogonal to the method’s motivation. Yet Sec. 2.2 motivates training because agents “prioritize task completion even when risk is implicit in the environment” (lines 125–128). This leaves the gap between the motivating problem and the demonstrated contribution unresolved. I therefore maintain my assessment.

We thank the reviewer for raising these concerns and respectfully clarify the following points:

(a) We would respectfully like to clarify a misunderstanding regarding the annotations. We transparently report the classifier_model field, which refers specifically to the three-category classification introduced in Response to Weakness 1: Benchmark Scope and Action Safety Analysis. This additional analysis was conducted during rebuttal to address your question about action safety. It is independent of the original benchmark annotation described in Appendix B.2. The original tasks, target actions, and coordinates were human-verified. The additional action-role labels group existing examples for analysis and do not change the benchmark ground truth. Additionally, given the unsafe content and ethical considerations associated with the benchmark, we released the data during rebuttal, as permitted by the submission requirements.

(b) Our training builds on verl and we provide all necessary components for implementing our algorithm. The released code includes constraint construction, component rewards, span alignment, and operation-level advantage allocation. We release these components with verl as a dependency, without redistributing the entire upstream framework. The appendix also documents the data generation, training, and evaluation procedures in detail, including the prompts used. Together, the paper and code specify our method and provide the information needed to reproduce it. We have followed the submission’s reproducibility requirements and hope that the paper will not be penalized for omitting upstream framework code or details unnecessary. The full code and models will be publicly released after organizing the dataset and experimental materials, upon the camera-ready stage.

(c) Our motivation is to improve safety awareness through constraint-conditioned training. This directly corresponds to encouraging agents to consider environmental risks before proceeding with task execution. Our method addresses the motivating problem by training agents to relate operational constraints to their actions, without directly training on harmful webpages. Direct training on harmful environments does not by itself teach this relationship. It can instead encourage environment-specific refusal and lead to over-refusal. Our experiments demonstrate that constraint-based alignment improves safety on unseen harmful environments while preserving task performance. Environmental risk recognition remains relevant to our motivation, while our demonstrated contribution concerns how agents use constraints to guide execution and refusal. We respectfully ask that the reviewer not penalize the paper for not adopting another training approach outside its stated contributions.

Finally, we thank the reviewer for the careful comments. Our core contributions are a benchmark that exposes safety weaknesses in existing GUI agents and a safety-alignment method that generalizes to harmful environments without directly using those environments for training. We welcome further questions and hope the reviewer will reconsider the score in light of these clarifications and the additional results provided in our other responses.

Reviewer_kMZ7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces an approach for preventing actions in harmful web environments. Using the GUI-WEBMINEFIELD benchmark, put together by the authors, they observed that SoTA GUI agents are sensitive to the execution of unsafe actions in such harmful web environments. Given this observation, the proposed approach combines Constraint-Conditioned Optimization to define permissible actions via safety constraints and Operation Advantage Calibration that partitions the output into token-level signals based on the activity performed by output elements. The empirical evaluation carried out using GUI-R1-7B as the baseline and its adaptation to support the approach demonstrates interesting results.

Critical strengths
  • The paper tackles an interesting and timely problem, well-aligned with the scope of ICLR.

  • The empirical results are promising.

  • Novelty: The GUI-WEBMINEFIELD is an interesting contribution to the manuscript, which, along with the different categories, enables the evaluation of agents in navigating through and performing different tasks within these webpages. Given that no comparable benchmark exists, the authors’ contribution to the benchmark is considered novel.

  • Motivation: The paper investigates a timely and tangible problem that is strongly related to the trustworthiness of Agentic AI in the context of safe task execution on the web. Thus, its motivation is credible, and the development of an approach that can safeguard users from unsafe actions due to harmful webpages is well elaborated.

Critical weaknesses
  • Novelty: Concerning the approach’s novelty, its key components are quite simple (i.e., generate constraints and fine-tune the model using RLVR and GRPO-like decomposed reward) but somewhat effective. Nevertheless, the novelty of the approach is considered modest, as it primarily relies on integrating and slightly adapting existing concepts and approaches. The lack of a theoretical foundation further reduces the novelty of the approach.

  • Rigour: The key part of the approach that enables assessment of its rigour is appropriately presented in the manuscript and the Appendix. A few things, however, deserve a closer look. First, it is not entirely clear whether any validity check takes place during constrained-conditioned optimisation to ensure that the safety constraints derived actually correspond to the harmful action. If no check takes place, then it might accidentally work, or the constraint might be overly strict (or the opposite). Second, the operation advantage calibration stage assumes that the output of the agent can be decomposed into the constituent elements (so that appropriate reward can be assigned to each element). The extent to which this decomposition is generic or feasible is not sufficiently articulated, and should be elaborated further.

  • Rigour: Another related concern is the validation of the approach. No benchmark or code is made available with the submission, which does not enable validation of the approach by inspecting the code or the benchmark. Stating that “The benchmark and code will be made publicly available” is not sufficient and does not enhance reproducibility.

  • Evaluation: The evaluation of the approach is detailed using a strong baseline model and assessing the capabilities of the approach by analysing its benefits when using its components independently. The evaluation on both the benchmark devised by the authors and another set of benchmarks from the literature is positive. However, some numbers in the evaluation section are not corroborated or do not match the numbers reported in the tables (e.g., l.375 “a 131.44% increase in refusal rate”, l.404 “operation advantage calibration increases RR by 38.16%”). This set of discrepancies raises questions about the overall validation maturity of the work. The authors should carefully revisit those sections. (Minor: Table 4 - GUI-Act-Web/GR GUI-R1-7B is the top w/o safety prompts)

Questions for authors

(1) Has any validity check been carried out during constrained-conditioned optimisation to ensure that the generated constraints are credible?

(2) Is there any theoretical foundation that can be used to support the soundness and rigour of the approach?

(3) Are there any concrete observations regarding the weaknesses of the approach for high-level GUI tasks? Stating that “given the safety requirements for GUI agents, this minor performance cost is generally acceptable” is really not acceptable, as this statement diminishes your work.

(4) Is there any reason why you have not provided the code and the benchmark at submission time?

2 comments

We thank the reviewer for the detailed assessment and helpful suggestions.

Response to Weakness 1 & Question 2: Theoretical Foundation and Novelty of the Approach

Theoretical Foundation. For a fixed rollout group at θ=θold\theta=\theta_{\mathrm{old}}, freeze tokens, spans, advantages and length weights. Compare only disjoint action/argument spans, with constant format rewards and KL and reasoning excluded. Define Gd(k)=wk∑t∈Sd(k)∇θlog⁡πθ(yt(k)∣x,y<t(k))G_d^{(k)}=w_k\sum_{t\in S_d^{(k)}}\nabla_\theta\log\pi_\theta(y_t^{(k)}\mid x,y_{<t}^{(k)}), where wk>0w_k>0 incorporates the same token-averaging weights in both methods. Let ARA_R denote the normalized aggregate reward R=λara+λprpR=\lambda_a r_a+\lambda_p r_p with positive weights, and Aa,ApA_a,A_p denote the separately normalized component rewards. The respective ascent-gradient contributions are

gGRPO=1K∑kAR(k)(Ga(k)+Gp(k)),g_{\mathrm{GRPO}}=\frac1K\sum_k A_R^{(k)}\left(G_a^{(k)}+G_p^{(k)}\right),

gOAC=1K∑k(Aa(k)Ga(k)+Ap(k)Gp(k)).g_{\mathrm{OAC}}=\frac1K\sum_k\left(A_a^{(k)}G_a^{(k)}+A_p^{(k)}G_p^{(k)}\right).

To analyze the gradient effect in a simplified setting, we model action and argument as binary variables, a∼Bernoulli(α)a\sim\mathrm{Bernoulli}(\alpha) and p∣a∼Bernoulli(βa)p\mid a\sim\mathrm{Bernoulli}(\beta_a), where α=σ(uθ)\alpha=\sigma(u_\theta) and βa=σ(va,θ)\beta_a=\sigma(v_{a,\theta}). All differentiable logits share parameters θ\theta. With rewards ra=a,rp=pr_a=a,r_p=p and unit reward weights, expected reward is J=α+(1−α)β0+αβ1J=\alpha+(1-\alpha)\beta_0+\alpha\beta_1.

At α=β0=β1=1/2\alpha=\beta_0=\beta_1=1/2, condition on two equal-length samples (1,0),(0,1)(1,0),(0,1) with averaging weight ww. Aggregate advantages are zero; component advantages are (c,−c)(c,-c) and (−c,c)(-c,c), c>0c>0, using the same normalization convention. Write U=∇θuθU=\nabla_\theta u_\theta and Va=∇θva,θV_a=\nabla_\theta v_{a,\theta} at this point. The Bernoulli score identity gives

g_{\mathrm{OAC}}=\frac{wc}{2}U+\frac{wc}{4}(V_0+V_1)=2wc\,\nabla_\theta J.$$ Thus, whenever $\nabla_\theta J\ne0$, a sufficiently small OAC ascent step increases expected reward. The result is conditional on this sampled group and symmetric probability point, where the two conditional argument distributions coincide but their parameter gradients may differ. It establishes a beneficial shared-parameter case, and general effectiveness is assessed empirically. Our contribution combines CCO with field-specific advantages and aggregate reasoning feedback. We detail this analysis in the revised paper. >**Response to Weakness 2 & Question 1: Constraint Validity and the Generality of Operation Decomposition** **Constraint Validity.** All synthetic constraints (1,000-example training set) were **manually inspected before use**. Any constraint that was ambiguous, inconsistent with the task, or failed to induce the intended aligned/conflicting relation was manually corrected or excluded. We clarify this validation step in the revised manuscript. Table 20 (revised paper) and Table 29 further show that CCO generalizes across safety prompts of varying strength and diversity. **Generality of Operation Decomposition.** Our method applies to structured GUI-agent outputs with distinct action, argument, and reasoning components. OAC can also extend to other RL tasks with structured outputs and multiple reward objectives, although **our experiments focus on GUI agents and their safety**. Table 25 shows that OAC improves task performance over GRPO without safety requirements, while providing greater gains in safety and usability under CCO. > **Response to Weakness 3 & Question 4: Releasing the Benchmark and Code for Reproducibility** We include the benchmark and code in the supplementary material and *[anonymous repository](https://anonymous.4open.science/r/GUI-WEBMINEFIELD-3E3F/)* . > **Response to Weakness 4: Clarifying Relative Improvements and Correcting Table 4** We appreciate the reviewer’s careful comments. The reported figures referred to relative improvements (131.44% and 38.16%). We revised the corresponding statements to report absolute improvements, and corrected the highlighting in Table 4. > **Response to Question 3: Limitations and Error Analysis on High-Level GUI Tasks** We thank the reviewer for pointing this out and apologize for the unclear wording. We observe inconsistent improvements across high-level benchmarks. We attribute this primarily to our method providing direct feedback on operation correctness but only indirect supervision on which subgoal to pursue next. For example, a GUI-Odyssey task asks the agent to “Use social media to post the feeling today and share it to a communication app.” Such broad goals require the model to infer intermediate steps and anticipate interface changes. Our safety constraints, however, are explicit and direct, deliberately avoiding ambiguous requirements. **This difference may introduce some interference with high-level task decisions during alignment training.** Table 5 further suggests that the inconsistent performance changes arise primarily from introducing CCO.

Dear Reviewer kMZ7,

We sincerely appreciate your positive assessment of our work and constructive suggestions. In response to your thoughtful feedback, our rebuttal provides clarifications and experimental analyses addressing the following key points:

  • Theoretical Foundation: We provide a theoretical analysis illustrating when OAC yields beneficial updates compared with aggregate-reward optimization.
  • Constraint Validity and Operation Decomposition: We clarify the manual validation of training constraints and the applicability of operation decomposition to structured agent outputs.
  • Reproducibility: We provide the benchmark and code in the supplementary material and anonymous repository.
  • Experimental Analysis and Corrections: We explain the inconsistent gains on high-level tasks, clarify the reported improvements, and correct the highlighting in Table 4.

We hope that these clarifications address your concerns and you would support our work. Thank you again for your valuable feedback and for considering our response during the rebuttal process. We will also address any further issues raised during the remaining discussion and incorporate the corresponding revisions into the camera-ready manuscript.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.