acceptodds
Under review as a conference paper at ICLR 2027

TidyUp: Post-Alignment of Agentic Layout Outputs with a Tiny Specialist

Abstract

Large language and vision–language agents can now assemble and edit real graphic designs, but they reason about geometry imprecisely: leaving overlaps, broken alignments, uneven spacing, and off-canvas content. The prevailing fix—re-invoking the same or a larger model to critique a rendered canvas and rewrite coordinates—is slow, costly, non-deterministic, and prone to regressing already-correct parts of the design. We argue that this narrow, geometric defect class does not call for more generalist reasoning but for a specialist. We present TidyUp, a tiny (86M-parameter), vision-free pipeline that repairs a design in place after the agent has already done the visual reasoning. A lightweight reflection router first flags which elements are broken and by what violation. Specialist correctors — first to jointly condition intricate design group structures and prior design state — apply hierarchical attention over local and global alignment geometry to nudge flagged elements into place, leaving size, type, and unedited elements untouched, with no regeneration. Unlike prior flow-based layout models, trained on noise-to-data, each corrector is trained for refinement, integrating its ODE from the broken layout. Owing to varying repair severity, TidyUp iterates reflection-correction with an auto-stop. Correctness-preserving guards—including a violation-regression bailout—guaranties TidyUp never returns a layout worse than its input. TidyUp, trained on AI-edited designs from a large-scale production system and a public Crello benchmark, with synthetic defects derived from multi-turn editing patterns, matches or exceeds the SOTA-level performance at a fraction of the cost: 1.2s latency at 250 RPM on a single H100, 160 MB footprint, with no additional agent calls.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
conditional flow, discrete diffusion, agentic editing

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_9rnoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TidyUp, a lightweight framework for correcting geometric errors in layouts produced by design agents. It uses a reflection router to identify overlap and alignment violations and two specialized flow-matching correctors to adjust element positions without changing their content or other attributes. The framework also incorporates group-aware attention, edit-aware conditioning, and iterative correction with regression checks. Experiments on an internal design dataset and the public Crello benchmark show improvements over existing layout generation and self-correction baselines, particularly in alignment quality and inference efficiency.

Critical strengths
  1. Practical and efficient layout correction. The paper addresses a meaningful problem in agentic design by separating geometric correction from expensive general-purpose reasoning. The proposed lightweight pipeline can repair layouts without rendering or regenerating content, making it potentially useful for multi-turn design workflows.

  2. Promising results and reasonably comprehensive ablations. TidyUp improves the reported gated quality scores on both datasets, with particularly noticeable gains in alignment correction. The ablation study also helps explain the contributions of the reflection router, refinement masks, group-aware attention, and edit-aware conditioning.

Critical weaknesses
  1. Inconsistencies in the evaluation metrics. The definition of the Gated Alignment Quality Score appears inconsistent across Appendix A.4.4 and Table 4. The text describes an average of Correction Rate and Preservation Rate, while the equations give different formulations. For example, the final configuration in Table 1 reports Correction Rate = 50.4 and Preservation Rate = 78.4, but the gated score is 63.6 rather than their average of 64.4. Similar discrepancies appear in the overlap scores. Since these metrics are central to the experimental comparisons, the calculation procedure needs to be clarified.

  2. Limited validation of the do-no-harm guarantee and real-world generalization. The guarantee in Section 4.6 relies on the violation count predicted by the router, which may not reflect all actual geometric errors or unintended element movements. Furthermore, the main quantitative experiments use synthetically corrupted layouts, while real multi-turn agentic editing is mainly demonstrated through qualitative examples. More direct evidence is needed to establish the reliability of the complete pipeline in realistic settings.

Questions for authors
  1. I noticed that the Gated Alignment Quality Score has different definitions in Appendix A.4.4 and Table 4. For example, the final model in Table 1 reports a Correction Rate of 50.4 and a Preservation Rate of 78.4, but the gated score is 63.6 rather than 64.4. Could the authors clarify which formula was actually used and how these scores were calculated?

  2. The paper claims that TidyUp never returns a layout worse than its input, but the safeguard appears to rely on the router's predicted violation count. What happens if the router misses an error or if a correction introduces a new problem? It would be helpful to see how often such cases occur in practice.

  3. Most of the quantitative results are based on synthetic layout defects. I wonder how well TidyUp performs on actual multi-turn agent-generated layouts. Could the authors provide a quantitative comparison on real editing outputs to support the claimed practical effectiveness?

  4. Since the main task is geometric layout correction, I wonder how TidyUp compares with optimization-based methods such as LayoutRectifier. Would a simpler constrained optimization approach achieve similar correction and preservation results?

1 comment

1. I noticed that the Gated Alignment Quality Score has ...

Thank you for identifying this inconsistency. Point 1 of our general response provides the exact formulas and explains this numerical example. The Gated Alignment Score averages Correction and Preservation over documents containing alignment defects, whereas the displayed Preservation column averages over all documents. Thus, averaging the two displayed columns does not reproduce the gated score. We have corrected the outdated appendix definitions and clarified the aggregation; reported values and rankings remain unchanged.

2. The paper claims that TidyUp never returns a layout worse than its input ...

We agree that Eq. 10 guarantees only a non-increasing router-estimated violation count, not preservation of true geometric constraints or design intent. We have revised the abstract and Section 4.6 accordingly; Point 2 of our general response details this clarification.

Missed errors. Figure 14(a) shows that the router classifies 19.9% of defective elements as unbroken, potentially leaving their defects unresolved. The routing masks guide rather than strictly constrain corrector movement, so unflagged elements may also be adjusted, occasionally repairing missed defects. This is not guaranteed, however additional Figure 16(b) shows the more than 50% of the missed but broken elements are indeed displaced towards fixing.

New problems. A proposed update is rejected if the router-estimated count increases. However, a new problem can pass the safeguard if the router misses it, or if other repairs offset it in the count. We therefore acknowledge this residual failure mode.

Figures 14(a–b), 16(a), Section A.6.5, quantify detection errors and unnecessary movement on valid designs. Point 3 of our general response and Table 5, Section A.7.1, report the complementary human evaluation, which identified one worsened design among 72 evaluated inputs. This is an end-to-end design-level regression rate, not a rate of new violations conditioned on accepted updates.

3. Most of the quantitative results are based on synthetic layout defects ...

Synthetic versus real defects. Figure 5 compares synthetic perturbations with real multi-turn agentic edits. Our induced edits more closely match the observed displacement frequency and magnitude, including the heavy tail, than the Gaussian perturbations used in prior work. Training covers a balanced range of defect severities. This supports the realism of the training distribution, but does not alone establish performance on natural errors.

We therefore added a blinded evaluation on 72 naturally defective designs produced by our pre-launch multi-turn workflow. Three independent judges with design expertise preferred TidyUp over VASCAR in 73.6%, 70.8%, and 75.0% of cases, respectively, and identified one design worsened by TidyUp. Point 3 of our general response and Table 5, Section A.7.1, provide the complete results.

4. I wonder how TidyUp compares with optimization-based methods such as LayoutRectifier...

We agree that constrained optimization is a relevant alternative and discuss it in Related Work. As noted in Point 4 of our general response, we could not locate a public LayoutRectifier implementation, and the available pseudocode lacked sufficient detail for a faithful reproduction and fair quantitative comparison.

Our choice of a learned repair pipeline reflects several practical differences in setting:

  1. Reference layouts at inference. LayoutRectifier retrieves training-set layouts to derive a grid system. TidyUp requires no retrieved reference layout or dataset access at inference.
  2. Overlap rules. LayoutRectifier specifies permissible overlaps through layout-type-specific child, parent, and other element sets. TidyUp uses a learned router to detect defects from the design context.
  3. Groups and edit history. LayoutRectifier represents structure through alignment and containment relations. TidyUp explicitly incorporates user-group hierarchies through group-aware attention and per-element edit-state conditioning.

These differences motivate TidyUp’s application setting; they do not establish superiority over optimization or preclude extending an optimizer with similar capabilities. Without a matched comparison, we cannot determine whether a simpler constrained optimizer would achieve comparable correction and preservation. We acknowledge this evaluation gap and present TidyUp as a complementary, reference-free, position-only learned approach.

Reviewer_m8nzOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

TidyUp is a geometric refinement system that operates after an LLM completes the design. It uses a Router to identify local overlaps and misalignments, and then invokes two lightweight flow-matching experts to correct object positions.

Critical strengths

The presentation is good.

The method has been evaluated and proved to be effective.

Critical weaknesses
  1. The paper does not systematically determine when TidyUp should intervene or establish a general recoverable error threshold, making it unclear how robust the method is to errors of different severities.
  2. TidyUp only modifies element positions (x, y) and cannot adjust size, font, text wrapping, or cropping, so it may resolve overlaps without addressing their underlying causes.
  3. TidyUp is trained and evaluated mainly on synthetically perturbed clean layouts, which may not reflect errors in real LLM-generated designs. Moreover, the comparison with general-purpose LLMs may not be entirely fair, since TidyUp is specifically trained for this repair setting.
  4. The paper does not adequately evaluate whether TidyUp may unnecessarily modify already well-designed layouts, which is important for real-world deployment, especially since many designs generated by large models may already be good enough.
  5. It is difficult to determine whether a correction is truly necessary in normal cases, as the evaluation mainly focuses on deliberately perturbed samples.
Questions for authors

pls see the weakness part.

3 comments

1. The paper does not systematically determine when TidyUp should ...

We agree that a formal recoverable-error threshold has not been established. We provide empirical evidence of how TidyUp’s behavior varies with severity:

  1. Severity coverage. Figure 5 (Appendix A.1) shows that our synthetic displacement distribution closely matches real multi-turn agentic edits, including the heavy tail of large displacements. Both the router and correctors are trained on a balanced distribution of severity levels, ranging from no displacement to small and large displacements.

  2. Effectiveness across defect severities (new Figure 15, Section A.6.5). We report the percentage of element defects repaired across severity levels for both defect types. Correction rates decline gradually with increasing severity, from 89% to 56% for overlap and from 60% to 25% for alignment, with no abrupt failure point observed. Overlap correction remains at or above 81% for d≤0.4d \leq 0.4, the range containing most evaluated elements, as shown in Figure 15.

  3. Router errors (new Figure 14(a)). We evaluate designs containing no defects, overlap defects, alignment defects, or both, across varying severity levels. The confusion matrix shows that 19.9% of defective elements are incorrectly classified as unbroken, while 18.4% of unbroken elements are incorrectly flagged as defective.

  4. Selective correction (new Figure 14(b)). Although 18.4% of elements in valid designs are flagged as problematic, only 6.6% of all evaluated elements are displaced beyond tolerance τ\tau. Conditioning on the design context allows the correctors to leave already well-placed elements unchanged, so router false alarms do not necessarily result in movement. Details are provided in Section A.6.5.

  5. Pre-production evaluation. Below, we also present results on designs produced by our pre-production multi-turn agentic workflow.

Together, these results indicate conservative behavior on valid designs and gradual performance degradation with increasing severity. We do not claim a formal intervention threshold; the revision explicitly documents the reduced correction rates at the largest severities (Figure 15).

2. TidyUp only modifies element positions (x, y) and cannot ...

We agree that some overlaps originate from inappropriate sizing, text wrapping, or cropping and cannot be fully resolved through translation alone. TidyUp targets placement errors in otherwise useful designs, rather than diagnosing every underlying cause or performing unrestricted design revision. Our contribution is to demonstrate that compact, task-trained models can repair local geometric defects without invoking a large reasoning model or regenerating the design.

Restricting edits to positions preserves the agent’s typography, element dimensions, and crop choices. The trade-off is that defects requiring changes to those attributes may remain unresolved, even if their visible impact is reduced. TidyUp therefore complements broader design reasoning rather than replacing it. We have clarified this limitation and identified extending repair beyond translation as future work.

3. TidyUp is trained and evaluated mainly on ...

Synthetic versus real defects. Figure 5 compares synthetic perturbations with real multi-turn agentic edits. Our induced edits more closely match the observed displacement frequency and magnitude, including the heavy tail, than the Gaussian perturbations used in prior work. Training covers a balanced range of defect severities. This supports the realism of the training distribution, but does not alone establish performance on natural errors.

We therefore added a blinded evaluation on 72 naturally defective designs produced by our pre-launch multi-turn workflow. Three independent judges with design expertise preferred TidyUp over VASCAR in 73.6%, 70.8%, and 75.0% of cases, respectively, and identified one design worsened by TidyUp. Point 3 of our general response and Table 5, Section A.7.1, provide the complete results.

Fairness of the comparison. We agree that TidyUp’s repair-specific training should be explicit. We include both learned baselines and an agentic self-correction alternative: LayoutFlow++ and LaySPA are trained on the corresponding data splits using public implementations, with LayoutFlow++ specifically trained for refinement; VASCAR uses GPT-5 with its published design-reflection procedure. VASCAR is a natural agentic alternative and also outperformed LaySPA in Table 2.

The comparison asks whether a compact specialist offers a better practical repair-quality–cost trade-off, not whether its architecture is superior under identical training and supervision. Point 5 of our general response clarifies this distinction.

4. The paper does not adequately evaluate whether ...

We agree that unnecessary intervention is important for deployment. As summarized in Point 2 of our general response, we have added a direct evaluation on initially valid designs in Section A.6.5.

Direct evaluation on valid designs. Figure 14(b) shows that, although the router flags 18.4% of elements as problematic, only 6.6% of all evaluated elements are displaced beyond tolerance τ\tau. The figure breaks down these interventions by predicted defect type and corrector. Additionally Figure 16(a) shows the extent and spread of these displacements. Figure 14(a) provides the router’s complete confusion matrix. These results show that false-positive flags do not necessarily lead to movement, although unnecessary displacement is not eliminated.

Complementary evidence on defective designs. Table 2 reports UnbrokenDrift for previously unchanged elements. TidyUp’s drift is lower than VASCAR’s on both datasets—2.15 versus 11.7 internally, and 0.61 versus 16.4 on Crello—while achieving the highest gated scores. This supports a favorable repair–preservation trade-off, rather than demonstrating minimal movement alone. The human evaluation in Section A.7.1, summarized in Point 3 of our general response, provides additional evidence on natural agent outputs.

Finally, when the router flags no defects, the correction loop terminates immediately. False alarms can nevertheless trigger unnecessary edits, so we acknowledge this residual deployment risk rather than claim that every valid layout remains unchanged.

5. It is difficult to determine whether a correction ...

We agree that evaluation on deliberately perturbed layouts alone cannot establish whether correction is necessary in ordinary use. We therefore added complementary evaluations on initially valid designs (Figures 14(a–b), Section A.6.5; see Point 2 of our general response) and naturally defective agent outputs assessed by blinded human judges (Table 5, Section A.7.1; see Point 3).

The former measures unnecessary intervention, while the latter evaluates the quality of repairs on real agent errors. These additions broaden the evidence beyond synthetic corruptions, but do not establish that every router-triggered correction is necessary. Assessing intervention necessity on an unfiltered mixture of valid and defective agent outputs remains a limitation.

Dear Reviewer,

Thank you for your time and constructive feedback. We have provided point-by-point responses, additional evidence, and corresponding manuscript revisions. We would appreciate your consideration of these updates and any reassessment you find warranted. Please let us know if further clarification is needed.

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

TidyUp is a compact, vision-free pipeline for repairing geometric defects in agent-edited graphic layouts. A learned router identifies overlap and alignment violations and dispatches them to two specialized flow-based correctors. The correctors modify positions while preserving element sizes and other attributes, using group-aware attention and optional edit-history conditioning. An iterative detection–correction loop rejects updates that increase the router’s predicted violation count. Training and quantitative evaluation use synthetically corrupted layouts from an internal production-design corpus and Crello. The paper reports improvements over layout-generation and self-correction baselines, with approximately 1.2 seconds latency per design.

Critical strengths

The paper addresses a practical problem: repairing local geometric defects without regenerating an otherwise useful design. Restricting corrections to position changes provides a clear architectural invariant that preserves element sizes, categories, and stacking order. The combination of specialized correctors, selective routing, and group-aware conditioning is well motivated for structured designs, and the ablations examine their contributions beyond refinement training alone. Evaluation on both an internal corpus and the public Crello dataset, together with reported low inference latency, makes the approach potentially useful as a lightweight component in agentic design workflows.

Critical weaknesses

First, the claimed “do-no-harm” guarantee is stronger than Equation 10 establishes: the acceptance rule only prevents an increase in the learned router’s predicted violation count. It does not guarantee preservation of true geometric constraints, violation severity, or user intent, particularly because unflagged elements may move and the router can miss errors. Second, the principal evaluation metrics are inconsistently specified. Appendix A.4.4 describes alignment quality as the mean of Correction and Preservation, its displayed equation instead uses Positional Quality—which the preceding text excludes for alignment—and Table 4 gives a different multiplicative formula involving insufficiently defined terms. The overlap results also do not match the stated arithmetic mean: Table 1’s base row reports Positional Quality 4.7 and Preservation 94.9, implying 49.8 rather than the reported 48.0. These discrepancies affect the primary evidence supporting the comparative claims.

Questions for authors
  1. Can the authors restrict the guarantee to non-increasing predicted violation count, or identify additional guards that establish a stronger property? What fraction of accepted corrections worsen independently measured constraints or human-assessed design quality despite passing the router-based bailout?

  2. Which exact formulas produced Tables 1 and 2? Please reconcile Appendix A.4.4 with Table 4, define Satisfaction Rate and its weighting parameter, and explain why the reported overlap scores differ from the stated arithmetic mean. Do the rankings remain unchanged under the corrected definitions?

  3. What quantitative evidence establishes repair quality on naturally occurring agent errors, beyond synthetic corruptions and qualitative examples? In particular, how often does TidyUp modify a legitimate placement or intentional overlap that an independent evaluator considers correct?

  4. How does TidyUp compare with optimization-based repair, such as LayoutRectifier, under matched position-only constraints?

2 comments

1. "Can the authors restrict the guarantee to non-increasing predicted violation count ..."

We agree that Eq. 10 guarantees only a non-increasing router-estimated violation count, not preservation of true geometric constraints or design intent. We have revised the abstract and Section 4.6 to make this distinction explicit and describe do-no-harm as a design objective. The position-only invariant preserves individual element attributes, but does not guarantee preservation of relationships between elements or overall design quality.

To assess unnecessary intervention, we added an evaluation on initially valid designs (Section A.6.5). The router flags 18.4% of elements, but 6.6% of all evaluated elements are displaced beyond tolerance τ\tau: 1.8% by the overlap corrector and 4.8% by the alignment corrector. Thus, not every false-positive flag results in movement; nevertheless, unnecessary movement remains a limitation. Figure 14(a) additionally reports the router’s class-wise confusion matrix.

We also added a blinded human evaluation on 72 naturally defective agent outputs (Section A.7.1). Judges identified one case in which TidyUp worsened the original design, corresponding to 1.4% of evaluated designs. This is an end-to-end, design-level regression rate, not a rate conditioned on passing the router-based bailout. We therefore do not interpret it as establishing a stronger guarantee.

2. "Which exact formulas produced Tables 1 and 2? Please reconcile Appendix A.4.4 with Table 4 ..."

Thank you for identifying these inconsistencies. Tables 1 and 2 were computed using the final Gated Score formulas provided in our general response. Appendix A.4.4 and Table 4 inadvertently retained an earlier formulation involving Satisfaction Rate and a weighting parameter; neither is used in the final scores. We have corrected the appendix and clarified the notation.

The apparent arithmetic discrepancy arises from different document subsets. Each Gated Score averages its components over documents containing the corresponding defect, whereas the separately reported Preservation column averages over all evaluation documents. Consequently, averaging the displayed component columns does not generally reproduce the Gated Score. The general response provides the exact equations and numerical examples addressing the reported discrepancies.

These revisions correct the metric documentation, not the underlying calculations: all reported values and rankings remain unchanged.

3. "What quantitative evidence establishes repair quality on naturally occurring agent errors ..."

Please see Point 3 of our general response and Table 5 (Section A.7.1) for the new blinded human evaluation on naturally defective agent outputs. It reports comparative repair preferences and human-assessed worsening relative to the original input.

For unnecessary intervention on initially valid designs, please see Point 2 and Figures 14(a–b), Section A.6.5. These complementary evaluations measure design-level degradation and element-level movement, respectively; they do not separately establish a human-validated modification rate for legitimate placements or intentional overlaps.

4. "How does TidyUp compare with optimization-based repair, such as LayoutRectifier, under matched position-only constraints?"

We have not conducted a matched position-only comparison with LayoutRectifier and therefore do not claim superiority over optimization-based repair. Point 4 of our general response explains the implementation and reproduction limitations, along with the differences in problem setting. We acknowledge this missing comparison and present TidyUp as a complementary learned approach supporting user groups and edit-state conditioning.

Dear Reviewer,

Thank you for your time and constructive feedback. We have provided point-by-point responses, additional evidence, and corresponding manuscript revisions. We would appreciate your consideration of these updates and any reassessment you find warranted. Please let us know if further clarification is needed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Response: Clarifications and Additional Evaluation

We thank the reviewers for the careful and constructive feedback. Their comments fall into five themes, and we have addressed each in the revision.

  1. Metric definitions (XNGC, 9rno).

We thank the reviewers for identifying these inconsistencies. Tables 1 and 2 were computed using our final metric implementation, but the equation in Appendix A.4.4 and Table 4 retained earlier formulations. We have corrected these definitions and clarified the aggregation procedure. No reported values or rankings change.

Let D\mathcal{D} denote all evaluation documents, and Dov\mathcal{D_{ov}} and Dal\mathcal{D_{al}} the subsets containing at least one overlap or alignment defect, respectively. These subsets are determined by ground-truth annotations and are identical across methods. Let PQdPQ_d, CorrdCorr_d and defect-specific PresdPres_d be positional-quality, correction rate and preservation rate for each document-level, then the final gated scores are:

• Gated Overlap Score = 1∣Dov∣∑d∈Dov12(PQd+Presd)\frac{1} {|D_{ov}|} \sum_{d \in D_{ov}} \frac{1}{2} (PQ_d + Pres_d)

• Gated Alignment Score = 1∣Dal∣∑d∈Dal12(Corrd+Presd)\frac{1} {|D_{al}|} \sum_{d \in D_{al}} \frac{1}{2} (Corr_d + Pres_d)

The discrepancy arises because the separately reported component columns use different evaluation subsets: Correction and Positional Quality are averaged over documents containing the corresponding defect, whereas Preservation is averaged over all documents. Within each Gated Score, however, Preservation is averaged over the corresponding defect-containing subset. Consequently, the Gated Score need not equal the arithmetic mean of the displayed component columns.

For the reviewers’ examples, the reported scores imply defect-subset Preservation of approximately 91.3 for the overlap base row and 76.8 for the final alignment configuration, compared with all-document Preservation of 94.9 and 78.4. We have made this distinction explicit in the metric definitions.

• Overlap, base row: 2 × 48.0 − 4.7 = 91.3

• Alignment, full model: 2 × 63.6 − 50.4 = 76.8

  1. Scope of “do-no-harm” (XNGC, 9rno).

We agree that Eq. 10 guarantees only a non-increasing router-estimated violation count, not preservation of true geometric quality or user intent. We have revised the abstract and Section 4.6 accordingly, describing do-no-harm as a design objective rather than an unconditional guarantee. Section A.6.5 now reports class-wise recall, including the unbroken class, and evaluates unnecessary intervention on initially valid designs: 6.6% of unbroken elements were displaced. This measures unintended movement, not necessarily perceptual degradation, and does not establish a stronger guarantee.

3. Evidence beyond synthetic defects (all reviewers).

Figure 5 already supports the realism of our synthesis by comparing displacement frequency and magnitude with real multi-turn edits, including their heavy tails. To complement the displacement-distribution evidence, Section A.7.1 adds a blinded evaluation on 72 defective designs produced during internal testing of a pre-launch multi-turn agentic workflow. Three independent judges with design expertise compared TidyUp and VASCAR corrections, preferring TidyUp in 70.8–75.0% of cases, with unanimous preference on 51 designs. The judges identified one case in which TidyUp worsened the original input.

4. Optimization-based repair (XNGC, 9rno).

We agree that a matched optimization baseline would strengthen the evaluation. We could not locate a public LayoutRectifier implementation, and the available pseudocode did not provide sufficient detail for a faithful reproduction. As described, its setting also differs: it retrieves training-set reference layouts at inference, optimizes sizes and positions, and uses hand-specified overlap rules. TidyUp instead targets practical agentic editing workflows through reference-free, position-only repair that explicitly models user groups and learns edit-state conditioning. These capabilities motivate TidyUp’s deployment setting, rather than establish superiority over optimization. We therefore present TidyUp as a complementary learned alternative and acknowledge the missing direct comparison.

5. Scope and fairness (m8nz).

TidyUp deliberately changes positions only, preserving element sizes, fonts, and crops. It therefore cannot resolve defects requiring changes to those attributes; extending repair beyond translation is future work. LayoutFlow++ and LaySPA are trained on the corresponding splits using public code, while VASCAR follows its published self-correction procedure. All methods receive the same evaluation inputs, but their training and supervision differ. The comparison therefore evaluates practical repair alternatives rather than isolating the effect of architecture under identical repair-specific training.

All additional results and revised claims are included in the updated manuscript, with changes marked in blue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.