acceptodds
Under review as a conference paper at ICLR 2027

One Pixel Wider, a Different Click: How Image Preprocessing Destabilizes GUI Grounding

Abstract

Stable GUI grounding is essential for reliable interaction by computer-use agents. However, we find that changing a screenshot dimension by a single pixel can alter which UI element a model selects, with these changes occurring disproportionately at particular image sizes. Why can one pixel produce a different click? This paper identifies dynamic-resolution preprocessing, which maps raw screenshot dimensions to discrete visual token grids or tile layouts, as one source of this instability. We call the image sizes where these configurations change preprocessing transition points. Across eight models and three benchmarks, the average UI-element switching rate nearly doubles at these points relative to adjacent one-pixel changes. Controlled interventions confirm that these configuration changes contribute to the excess switching. Further analysis in two models shows that resizing and content alignment affect grounding beyond changes in token count alone. These findings motivate three investigations. First, because switching concentrates at transition points, we show how to “steal” hidden preprocessing settings from clicks. The inferred settings help predict unstable screenshot sizes, with predictions validated on unseen screenshots. Second, we use sensitivity to small movements of interface content to detect errors: clicks whose selected UI element changes under these movements are more likely to be incorrect. Combining this consistency check with model confidence improves error detection. Finally, guided by the effects of resizing and content alignment, we design a mitigation method that improves grounding stability. Together, these analyses and investigations offer insights into why one pixel can change a click and how to improve the stability of GUI grounding.

Primary area:
foundation or frontier models, including LLMs
Keywords:
GUI Grounding, Computer-Use Agents, Vision-Language Models, Image Preprocessing, Prediction Stability

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_HXZfOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies instability in GUI grounding models caused by one-pixel changes in screenshot dimensions. It identifies dynamic-resolution preprocessing transitions as a source of UI-element switching, validates the phenomenon across multiple models and benchmarks, proposes a black-box probe for inferring hidden preprocessing settings from click behavior, uses shift consistency for error detection, and introduces Grid-Phase Canonicalization (GridPC) as a lightweight mitigation.

Critical strengths

This is a careful and well-motivated study of a concrete failure mode in computer-use agents. The paper shows why average accuracy can hide instability and strengthens the diagnosis with transition-pair controls, cross-model tests, and controlled interventions. It also provides practical follow-ups: a black-box preprocessing probe, a useful consistency signal, and a lightweight mitigation that does not require training.

Critical weaknesses

The mitigation and reliability claims need stronger downstream validation. GridPC is mainly evaluated in controlled static settings, whereas real agents face browser scaling, scrolling, OS rendering differences, dynamic content, overlays, and multi-step feedback; evidence on interactive task success is missing. The switching analysis also depends on UI-element parsing and assignment rules, so more detailed robustness checks are needed to rule out parser errors or boundary choices. Finally, shift consistency requires extra forward passes and should be compared with multi-view aggregation or sampling-based uncertainty under matched compute.

Questions for authors
  1. How sensitive are the main switching conclusions to the UI parser and element-assignment rules?
  2. Does GridPC improve downstream task success in an interactive benchmark such as OSWorld?
  3. Under equal compute, how does shift consistency compare with multi-view or sampling-based uncertainty baselines?
7 comments
AuthorsOpenReview ↗

Response to Q1–Q2: robustness to element assignment and downstream relevance

We appreciate your constructive feedback and positive assessment of the study’s motivation and practical contributions. We hope the explanations below answer your questions. Please let us know if any points need further clarification.


Q1/W: How sensitive are the main switching conclusions to the UI parser and element-assignment rules?

Thank you for this question. The excess switching persists under human judgments and alternative element-assignment rules. Table 6 reports a blinded human evaluation of Qwen3-VL-8B on 300 screenshots, with 100 each from ScreenSpot-Pro, UI-Vision, and OSWorld-G. Each screenshot contributed one transition pair and two adjacent controls, giving 900 prediction pairs. One expert judged whether the clicks selected different UI elements without seeing the padding amounts, grid dimensions, or pair type. Excess switching is the transition switching rate minus the mean rate of the two adjacent controls. Human judgments yield 11.33 percentage points of excess switching (95% CI [7.50, 15.33]), compared with 11.17 points [7.33, 15.17] under automatic assignment.

During the rebuttal period, we reanalyzed the same human annotations using two alternative definitions of switching. Counting switches only between elements with different functions gives 21.33% switching at transitions and 10.00% for adjacent controls, an excess of 11.33 percentage points. Also counting changes between elements and background gives an excess of 14.67 percentage points. The transition effect persists under both definitions. We also compared three automatic assignment rules on the same 900 pairs, keeping the OmniParser detections and model predictions fixed:

Element-assignment ruleExcess switching (percentage points)95% CI
Paper rule: use the smallest containing interactive region, or the smallest containing region if none is interactive11.17[7.33, 15.17]
Smallest containing region, regardless of interactivity14.67[10.50, 18.83]
Smallest containing region, or the nearest region within 16 pixels for unmatched clicks15.17[10.83, 19.67]

The first two rules leave clicks outside all detected regions unassigned. All pairs remain in the denominator, and pairs with unassigned clicks do not count as switches. The table uses 10,000 bootstrap resamples of screenshots within each benchmark, with the same draws across rules.

All three rules yield positive excess switching, with 95% CIs above zero. The transition effect therefore persists under alternative assignment rules. We will add these robustness checks to the revised paper.


Q2/W: Does GridPC improve downstream task success in an interactive benchmark such as OSWorld?

Thank you for this valuable suggestion. During the rebuttal period, we added two experiments on MiniWoB to explore the potential value of stabilizing preprocessing for interactive tasks. Specifically, we examined how preprocessing affects multi-step task outcomes and how GridPC behaves at different rendering scales.

For multi-step outcomes, we tested Qwen3-VL-8B on eight MiniWoB task templates under two settings: an intervention at one random intermediate click, and repeated interventions at preselected random steps. Each setting used 696 task instances, with a maximum of 30 decisions per instance. We varied screenshot padding during grounding and executed the predicted clicks in the original browser, with planning and error recovery enabled. Compared with adjacent one-pixel controls that preserve the preprocessing configuration, transitions increased final-outcome switching by 6.03 percentage points (95% CI [4.24, 7.90]) for a single intervention and 4.38 points [2.80, 6.03] for repeated interventions. Final-outcome switching counts pairs with different success or failure outcomes. The intervals use bootstrap resampling stratified by task. Identical-input repeats showed zero switching. These results show that grounding changes caused by preprocessing can propagate to final task outcomes.

For rendering scale, we evaluated Qwen3-VL-8B and UI-Venus-1.5-8B on 64 MiniWoB scenes (173 targets) rendered in Chromium at DPR 1.0 and 1.25 with unchanged DOM/CSS layout. Within each DPR, GridPC maps variants differing by one pixel of padding to identical processed inputs and yields zero switching. This demonstrates its effectiveness against one-pixel padding at both rendering scales.

On the other hand, we would like to clarify the scope of our work. Our central contribution is to identify and causally characterize preprocessing as a structural source of GUI grounding instability. GridPC serves as a simple mitigation guided by these findings, and its evaluation focuses on grounding stability. Evaluating whether GridPC improves end-to-end task success on benchmarks such as OSWorld is a valuable direction for future work. We will include this discussion and these experiments in the revised paper.

AuthorsOpenReview ↗

Response to Q3: shift consistency under a matched inference budget

Q3/W: Under equal compute, how does shift consistency compare with multi-view or sampling-based uncertainty baselines?

Thank you for suggesting this comparison. During the rebuttal period, we compared shift consistency with sampling-based consistency on Qwen3-VL-8B and UI-Venus-1.5-8B. Each model was evaluated on the paper's 2,010 targets: 600 from ScreenSpot-Pro, 900 from UI-Vision, and 510 from OSWorld-G.

Both methods use four forward passes and share one greedy reference prediction. Shift consistency uses three additional deterministic shift probes. Sampling-based consistency uses three independent samples of the same image at temperature 1. We report AUROC for detecting incorrect reference predictions, where higher is better. We also combine each consistency signal with the same token confidence score by averaging their risk ranks, without fitting parameters or additional forward passes.

Error detection signalQwen3-VL-8B AUROCUI-Venus-1.5-8B AUROC
Sampling consistency0.7900.733
Shift consistency0.8030.752
Sampling consistency + token confidence0.85620.7973
Shift consistency + token confidence0.86260.8081

With the same number of forward passes, shift consistency achieves higher AUROC on both models, both as a standalone signal and when combined with token confidence. We will include these results and the comparison protocol in the revised paper.

Reviewer_HXZfOpenReview ↗

Response and maintained rating

Thank you for the comprehensive response and additional analyses. The alternative element-assignment rules and human judgments strengthen the core finding, and the matched-budget uncertainty comparison directly addresses one of my questions. The MiniWoB experiments also provide useful evidence that preprocessing-induced grounding changes can propagate to task outcomes.

My remaining concern is that the downstream mitigation claim is still not fully established. The new multi-step experiment demonstrates that preprocessing transitions affect final outcomes, but it does not directly show that GridPC improves end-to-end task success in a realistic interactive benchmark such as OSWorld. Its robustness beyond controlled padding, shifts, and selected rendering changes therefore remains uncertain. I will maintain my original rating of 2 (Weak Rejection).

Reviewer_kfzeOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies how one-pixel screenshot changes destabilize GUI grounding through dynamic-resolution preprocessing. Switching increases at transitions in the token grid or tile layout. Configuration-fixing and transition-shifting interventions isolate the preprocessing effect, while further experiments examine resize scale and grid phase. The paper also explores preprocessing-rule inference, consistency-based error detection and a training-free canonicalization method, GridPC.

Critical strengths
  1. The main finding is informative and practically relevant: similar mean accuracy can hide substantial changes in which UI element is selected. The matched adjacent controls, fixed-configuration interventions and shifted transition points provide a convincing causal design for attributing excess switching to preprocessing rather than merely observing generic input sensitivity.

  2. The breadth and mechanism analyses strengthen the result. The study covers eight models, four preprocessing families and three benchmarks, with screenshot-cluster bootstrap intervals and human validation. Separating resize scale, token count and grid phase gives a useful explanation of the effect. These findings are my main reason for leaning toward acceptance.

Critical weaknesses
  1. The practical scope remains uncertain. Repeated-edge padding and content shifts hold screenshot content fixed, whereas real window resizing and DPI changes can alter layout and rendering. The relatively high switching rate away from transitions also makes it important to distinguish the transition-specific effect from general fragility. The paper establishes the controlled phenomenon more strongly than its impact on multi-step agent outcomes.

  2. GridPC's near-zero switching is not sufficient evidence of broader robustness. The appendix reports that canonicalization makes the processed inputs identical for the evaluated padding steps and most shift steps. This invariance is useful, but its evaluation should separate identical-input cases from changes that survive canonicalization. The modest accuracy gains and absence of a simple fixed-resolution comparison make the application less convincing than the diagnostic study.

Questions for authors
  1. Can you separate GridPC results for identical and non-identical processed inputs, and compare with a fixed canonical resolution under the same evaluation protocol? How does it behave under non-replicated padding or actual content-scale changes such as 100% versus 125% DPI?

  2. Do the transition effects persist on ScreenSpot-Pro cases dominated by pixel-budget downscaling? Is there evidence that the induced grounding switches change multi-step task outcomes? The latter would clarify scope rather than being a prerequisite for the controlled causal finding.

  3. For shift-based error detection, how does performance compare with sampling-based consistency at the same number of forward passes? Please distinguish its standalone signal from combinations with confidence scores.

5 comments
AuthorsOpenReview ↗

Response to W1–W2: practical scope and GridPC evaluation

Thank you for the thoughtful review and for recognizing our causal design and mechanism analysis. We hope our responses address your concerns. We welcome any further questions.


W1: Scope of controlled perturbations, transition-specific instability, and downstream impact.

W1-1 Real window resizing and DPI changes

Thank you for highlighting this distinction. Our central contribution is to identify and causally characterize preprocessing as a structural source of GUI grounding instability. Real window resizing and DPI changes can jointly alter layout, rendering, and preprocessing. We use padding and content shifts to preserve scene content and target semantics, allowing us to isolate the effect of preprocessing. To further examine the effect of rendering scale, we additionally evaluate scenes rendered at device pixel ratios (DPR) of 1.0 and 1.25 with unchanged DOM/CSS layout (please see our response to Q1). GridPC preserves one-pixel padding invariance within each tested rendering scale.

We would also like to clarify the intended scope of our work. Our primary aim is to identify and explain preprocessing as a source of GUI grounding instability, with GridPC serving as an initial mitigation guided by these findings. These findings provide a useful foundation for future work on how preprocessing interacts with layout and rendering changes in real interfaces.

W1-2 Additional instability at preprocessing transitions

In Sec. 3.1 of the paper, we distinguish the additional instability caused by preprocessing transitions from switching that occurs without them. Excess switching is the difference between switching rates at transitions and adjacent one-pixel controls that preserve the token grid or tile layout (Fig. 3(a)). Fixing the configuration reduces this excess (Sec. 3.1 and App. B.3), while moving the transition points shifts the screenshot dimensions at which excess switching occurs (Table 2). These interventions support the causal role of preprocessing transitions.

W1-3 Multi-step task outcomes

During the rebuttal period, our added controlled multi-step experiments show that preprocessing changes can propagate from grounding decisions to final task outcomes (Q2). Multi-step task success additionally reflects planning and error recovery. Our controlled analysis isolates how preprocessing changes individual grounding decisions, using element switching and correctness flips to reveal instability that aggregate outcomes can obscure. In particular, changes from correct to incorrect and vice versa can offset in average accuracy.

We will clarify these points and include the new experiments in the revised paper.


W2: GridPC invariance, fixed-resolution comparisons, and accuracy.

W2-1 Identical inputs and broader robustness

Thank you for this suggestion. GridPC is a simple mitigation guided by our mechanism analysis, and making processed inputs identical is an intended part of its design. We report identical and non-identical cases separately for Qwen3-VL-8B and UI-Venus-1.5-8B on the screenshots described in Q1. In the original scans, 95.41% of pairs become identical, and the main stability gain comes from this canonicalization. For the remaining non-identical pairs, switching is 9.55% with native preprocessing versus 9.14% with GridPC, with a paired difference CI that includes zero.

On the same screenshots and models, we also tested non-replicated padding by mirroring the image content at the borders. Only 10.42% of pairs become identical. GridPC reduces transition switching from 15.36% with native preprocessing to 5.60%. Among pairs whose processed inputs remain different, switching decreases by 8.33 to 10.94 percentage points for both models under padding and content shifts, with all four paired 95% CIs excluding zero.

W2-2 Fixed-resolution comparison

During the rebuttal period, we added global and per-image fixed-resolution controls using the same screenshots, models, and evaluation protocol (please see our response to Q1). Averaged across models and padding/shift experiments, switching is 10.68% and 10.17%, respectively, versus 0.42% for GridPC. The global control uses a comparable visual token budget. Thus, fixing the output resolution alone does not reproduce the stability gained through canonicalization.

W2-3 Stability and accuracy

GridPC improves stability under padding and content shifts and maintains or modestly improves grounding accuracy. In Table 4, replacing native preprocessing with GridPC increases accuracy from 45.62% to 46.59% for Qwen3-VL and from 59.39% to 60.95% for UI-Venus-1.5 on the original screenshots. Stability and accuracy measure different properties: changes from correct to incorrect and vice versa can offset in average accuracy. GridPC reduces switching that average accuracy alone can obscure.

We will incorporate these analyses and comparisons in the revised document.

AuthorsOpenReview ↗

Response to Q1: GridPC evaluation and additional controls

Q1: GridPC results for identical and non-identical inputs, fixed-resolution controls, and additional perturbations.

Thank you for these questions. During the rebuttal period, we conducted the following additional analyses. The first three use the same 96 screenshots per model, evaluating Qwen3-VL-8B and UI-Venus-1.5-8B. Results equally weight screenshots, models, and padding/shift experiments. We compute 95% CIs using 10,000 bootstrap replicates clustered by screenshot.

Identical and non-identical processed inputs. We group pairs by whether their processed inputs are identical under GridPC and compare native preprocessing and GridPC on the same pairs from the original dense scans. GridPC makes 95.41% of pairs identical. Following the evaluation protocol in App. F, we reuse one greedy prediction for identical processed inputs, yielding zero switching for these pairs under GridPC. For pairs whose processed inputs remain non-identical under GridPC, switching is 9.55% (native) versus 9.14% (GridPC), with a paired difference CI that includes zero. The main gain on these scans therefore comes from mapping the tested variants to identical processed inputs.

Fixed resolution. Global fixed uses one output resolution for all screenshots, approximately matching the average visual token budget of GridPC. Per-image fixed uses the output resolution that GridPC assigns to each original screenshot and keeps it unchanged across all padding/shift variants. Both controls resize each input while preserving its aspect ratio, then center it with padding to reach the chosen output size. The table reports switching across the original dense scans, averaged over screenshots, models, and padding/shift experiments.

MethodSwitching (%)
Native9.76
Global fixed10.68
Per-image fixed10.17
GridPC0.42

Non-replicated padding. We test two ways to add padding: filling it with a constant gray value or mirroring the image content at the borders. Compared with native preprocessing, GridPC reduces transition switching from 14.84% to 0.13% with gray padding, and by 9.77 percentage points [95% CI: 6.51, 13.02] with the second method. For the second method, only 10.42% of pairs become identical under GridPC. Among pairs whose processed inputs remain different, switching decreases by 8.33 to 10.94 percentage points for both models under padding and content shifts. The paired 95% CI excludes zero for each of the four combinations of model and perturbation type.

Rendering scale. Changing DPR changes how text and icons are rendered. We render 64 MiniWoB scenes (173 targets) in Chromium at DPR 1.0 and 1.25 with unchanged DOM/CSS layout. Using Qwen3-VL-8B and UI-Venus-1.5-8B, we evaluate padding stability within each DPR and correctness changes across the two DPRs. Within each DPR, one-pixel padding produces identical GridPC inputs and zero switching, preserving the intended padding invariance at both tested rendering scales. Averaging both models, scenes and both DPRs, accuracy is 66.90% (native) versus 74.02% (GridPC). Across DPR 1.0 and 1.25, correctness flips are 19.22% with native preprocessing and 26.97% with GridPC. The paired difference CI includes zero.

AuthorsOpenReview ↗

Response to Q2–Q3: downstream effects and matched-budget comparisons

Q2: Pixel-budget downscaling and multi-step outcomes.

Great questions. Thank you for raising these points.

Pixel-budget downscaling. Yes, the transition effect persists on ScreenSpot-Pro screenshots subject to pixel-budget downscaling for Qwen3-VL. Using the same pixel budget as the paper, we evaluate all 81 ScreenSpot-Pro screenshots that require downscaling to meet the pixel budget for Qwen3-VL. Switching rises from 7.41% for adjacent controls to 13.58% at transitions, an excess of 6.17 percentage points [95% CI: 2.16, 10.49].

Multi-step outcomes. Yes, our added multi-step experiments show that grounding changes caused by preprocessing can propagate to final task outcomes, including changes from success to failure and vice versa. We test Qwen3-VL-8B on eight MiniWoB task templates (696 instances per setting, up to 30 decisions per instance). For grounding, we vary only screenshot padding and execute the predicted clicks in the original browser, with planning and recovery enabled. The table reports the increase in final-outcome switching at preprocessing transitions compared with adjacent one-pixel controls that preserve the preprocessing configuration. We compute 95% CIs using bootstrap resampling stratified by task.

InterventionIncrease in final-outcome switching (percentage points)95% CI
One random intermediate click6.03[4.24, 7.90]
Repeated interventions at clicks occurring at preselected random steps4.38[2.80, 6.03]

Identical-input repeats show zero switching. Among the 62 and 52 transition pairs with different outcomes, respectively, five also satisfy stricter trace criteria: the same state and target before intervention, different UI states after the click, and further actions in both branches.


Q3: Shift consistency versus sampling under matched inference budgets.

Thank you for suggesting this comparison. Under the same inference budget, shift consistency achieves modestly higher error detection AUROC than sampling-based consistency on both evaluated models, both as a standalone signal and when combined with token confidence.

During the rebuttal period, we compared both methods on 2,010 targets per model using four forward passes: a shared greedy reference prediction plus three deterministic shift probes or three independent samples of the same image (temperature 1).

The table reports error detection AUROC (higher is better). We combine each consistency signal with token confidence by averaging their risk ranks without fitting, with no additional forward passes.

ModelSignalShift consistencySampling-based consistency
Qwen3-VL-8BStandalone0.8030.790
Qwen3-VL-8BCombined with token confidence0.86260.8562
UI-Venus-1.5-8BStandalone0.7520.733
UI-Venus-1.5-8BCombined with token confidence0.80810.7973

We will add this comparison to the revised paper.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to Reviewer nvvV: GridPC generalization and preprocessing rule reliability

Thank you for your encouraging assessment of our causal findings and their practical applications. We hope our replies clarify the points you raised. Further questions and comments are always welcome.


Q1: Extending GridPC to tiled preprocessing and tile placement.

Thank you for this helpful suggestion. We see this mainly as an engineering challenge. Our InternVL3.5-8B results support alignment correction for tiled preprocessing (Table 14). The extension retains the thumbnail and uses dimensions divisible by 448, the tile size. We scan 96 screenshots (32 each from ScreenSpot-Pro, UI-Vision, and OSWorld-G) along both axes with replicated borders: padding from 0 to 64 pixels and content shifts from 0 to 32 pixels. Switching averages adjacent one-pixel steps within screenshots, then screenshots equally.

At the same maximum of 7 tiles, adding phase correction to pad-to-grid reduces content-shift switching from 20.75% to 1.25%. With a limit of 12 tiles, a separate paired native → GridPC comparison gives 14.49% → 0.20% for padding and 17.40% → 0.86% for content shifts. These are stability benefits, distinct from accuracy gains.

The thumbnail changes even when content tiles are preserved (App. C.5). A promising extension would estimate phase on the tile grid to choose padding offsets that keep salient content away from tile boundaries while preserving token alignment within tiles. This requires joint control of tile layout, scale, and thumbnail.


Q2: Compatibility with model training and geometric stress tests.

This is a great question. GridPC preserves scene content within the native pixel budget, downscaling uniformly only when needed. We test Qwen3-VL-8B and UI-Venus-1.5-8B without retraining on geometric subsets of 2,010 targets across 1,450 screenshots from these three benchmarks. Small UI elements are the bottom 10% by target box short side after native preprocessing (201 targets on 151 screenshots per model, 1.53–17.00 pixels). High aspect ratios are the top 10% of screenshots by max(W/H, H/W) (145 screenshots, 180 targets, ratios 1.905–3.556). Selection uses geometry alone and stays fixed across methods for each model.

Stress subsetAccuracy change (pp)Change on remaining inputs (pp)Switching (%) native → GridPC
Small UI elements+1.60 [−2.78, 5.95]+1.01 [−0.19, 2.20]9.78 → 0.49
High aspect ratios+1.55 [−3.16, 6.25]+1.23 [0.02, 2.41]13.79 → 0.63

Changes are GridPC minus native accuracy on original screenshots before perturbations. We average targets within screenshots, then screenshots and models equally. For each model, switching covers all 201 small UI elements on 151 screenshots and all 180 targets on 145 screenshots with high aspect ratios, using the same scanning protocol as Q1. Axes and scan types are equally weighted. Brackets give paired 95% CIs from 10,000 screenshot bootstrap resamples within benchmarks.

Both models show lower switching in both stress subsets, with all paired 95% CIs for the reductions excluding zero. Observed mean accuracy changes are positive and close to those on remaining inputs.


Q3: Reliability of inferred preprocessing rules under closed-weight model variability.

Thank you for highlighting this practical concern. Output variability can affect precise rule selection. We test whether inferred transitions predict additional instability despite this variability. For GPT-5.4 nano and GPT-5.6 Luna, App. D uses 150 discovery and 150 held-out screenshots, 50 per benchmark in each split. At temperature zero, we add 0 to 64 pixels of right or bottom padding in one-pixel steps. Identical-input repeats measure API variability. Rule stability measures how often repeating the search on bootstrap resamples of discovery screenshots recovers a rule.

With the rule fixed, held-out validation compares rates of click movements exceeding 1% of the original screenshot diagonal at predicted transitions versus other one-pixel steps under the same API settings. Rule frequencies and held-out 95% CIs use 10,000 screenshot resamples within benchmarks.

For nano, the 32-pixel ceiling rule has 67.60% rule stability and a held-out rate increase of 15.86 percentage points [10.85, 20.92]. For Luna, the 32-pixel floor and ceiling rules are selected in 40.97% and 27.68% of resamples, respectively. The selected floor rule gives an increase of 4.64 points [−0.04, 9.40]. The documentation comparison in App. D agrees with stride and rounding for nano, and stride for Luna.

Practitioners should repeat identical inputs, report bootstrap rule frequencies, and validate on held-out screenshots. Report confidence in stride and rounding separately. If rounding is ambiguous, report competing rules and assess them with additional repeats.

We will incorporate these suggestions and revisions into the revised paper.

Area_Chair_uxkNOpenReview ↗

Reviewers kfze and nvvV] Urgent Reminder: Discussion Period Ends November 18

Dear Reviewers kfze and nvvV,

The discussion period will end on November 18, 2026. Please carefully review the authors' responses and actively engage in discussions.

If you have not yet initiated a discussion, please do so immediately. Reviewers are expected to address the authors' rebuttals, provide constructive follow-up comments, and update their evaluations when appropriate.

Please also ensure that your reviews reflect independent human judgment and that any AI assistance is properly documented in accordance with the review guidelines.

Active participation is an essential reviewer responsibility and is critical to ensuring a rigorous and fair review process.

Thank you for your cooperation.

Reviewer_nvvVOpenReview ↗

Official Review of Submission2380 by Reviewer nvvV

Thank you for your clarification and additional analysis, which resolve my concerns well. I'll keep my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.