acceptodds
Under review as a conference paper at ICLR 2027

CADFlow: Recovering the Construction Process for Sketch-to-CAD Generation

Abstract

Generating accurate and executable parametric CAD models from raster sketch images remains challenging. Existing methods either rely on individually accessible strokes and learn to group them into CAD operations, requiring additional stroke recovery for raster sketch inputs, or directly generate CAD sequences from raster sketch images, where insufficient intermediate-state context increases the risk of erroneous construction steps and invalid geometries. To address this problem, we propose CADFlow, which formulates sketch-to-CAD generation as construction-process recovery. CADFlow employs CAD Construction Process Recovery (CCPR) to recursively recover preceding-state features from the final input state to an empty state. It then decodes adjacent-state transitions in the forward direction into a parametric CAD sequence. To alleviate error accumulation, CADFlow further employs Error-Aware Redrawing (EAR) to detect and locally redraw erroneous CAD operations and optimize sketch parameters under predicted geometric constraints. Experiments on DeepCAD and Fusion 360 demonstrate that CADFlow generates accurate, executable, and geometrically consistent CAD models while exhibiting strong cross-dataset generalization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
CAD Generation, Sketch-to-CAD, Construction Process Recovery, Error-Aware Redrawing, Geometric Constraint Refinement

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fkAWOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the generation of executable parametric CAD sequences from a single raster sketch. CAD Construction Process Recovery (CCPR) recursively predicts preceding construction-state features from the final image until reaching an empty-state reference, then decodes adjacent-state transitions into sketch-extrude operations in the forward direction. Error-Aware Redrawing (EAR) detects erroneous operations in the initial sequence, regenerates the affected operations, and refines line endpoints using GCN-predicted geometric relations and SLSQP. The model is jointly trained with five supervised losses while freezing the DINOv2 backbone. Ground-truth intermediate states are used for training supervision, not as inference inputs. Evaluation covers DeepCAD, zero-shot transfer to Fusion 360, hand-drawn sketches, component ablations, construction-prefix accuracy, computational cost, and failure cases.

Critical strengths
  1. Explicit state recovery has substantive, partially controlled experimental support. The method connects the ambiguity of composite raster sketches to operation-level state transitions and supervises recovery using features of actual construction states. In Table 3, the invalid ratios for Forward + state supervision, Backward without state supervision, and CCPR are 21.0%, 19.5%, and 4.81%, respectively. Table 5 reports the same parameter count and computation for these three recovery variants: 49.22M parameters and 33.61 GFLOPs. These comparisons support the combined use of backward recovery and state supervision under the reported setting; differences in parameter count or GFLOPs among these variants do not explain the improvement.

  2. The system evaluation covers reconstruction quality, transfer, refinement, and computational trade-offs. The paper reports in-domain evaluation, direct cross-dataset transfer, and hand-drawn inputs, with all task-specific training conducted on DeepCAD. CADFlow achieves 82.4% parameter accuracy and 3.9% invalid ratio on DeepCAD, and 77.8% parameter accuracy and 9.7% invalid ratio on Fusion 360. Compared with redrawing every operation, selective redrawing improves command accuracy from 78.6% to 82.3% while reducing median latency from 183.70 to 132.44 ms. The hand-drawn evaluation and explicit discussion of failures on invisible back-side details further clarify the practical scope of the system.

Critical weaknesses
  1. The evaluation protocol and baseline adaptations are not specified well enough to verify the central comparisons. Section 4.1 gives only high-level definitions of command accuracy, parameter accuracy, invalid ratio, and Chamfer distance. It does not fully specify parameter tolerances and denominators, CD point sampling and distance convention, normalization and aggregation, or how invalid outputs are handled when computing CD. The textual definition of invalid ratio refers to CAD sequences, whereas Tables 1 and 2 also compare B-rep and CADQuery outputs; a common geometric-validity criterion across these representations needs clarification. Appendix A.2 describes shared source data, retraining, and input-module adaptations, which are positive steps. However, the actual evaluation counts and conversion/filtering rules, baseline versions, feature connections and freezing choices, and training/tuning budgets remain insufficiently documented. Free2CAD also receives ordered stroke information unavailable to raster-only methods, so this difference in input conditions should be explicit. These omissions limit confidence that the reported ranking isolates the contribution of CADFlow. They establish a need for clarification, rather than demonstrating that the comparisons are unfair.

  2. The evidence does not fully validate the claimed mechanisms or their reliability boundaries. The matched recovery variants provide useful evidence, but direct generation and CCPR require 16.47 and 33.61 GFLOPs, respectively. This comparison does not fully separate the effect of construction-process modeling from differences in architecture, supervision, and computation. The EAR ablations support an overall benefit, but the paper does not report error-detection precision/recall, the frequency of changing correct operations into incorrect ones, or selection comparisons using the same redrawing model and initial predictions. Geometric constraint refinement also lacks direct measurements of relation-prediction quality, angular residuals, constraint satisfaction, and solver success or fallback rates. Adding geometric constraint refinement changes parameter accuracy from 82.3% to 82.4%, invalid ratio from 3.96% to 3.88%, and the scaled CD reported in Table 3 from 14.3 to 13.7, while increasing median latency from 132.44 to 173.15 ms. Without repeated-run variability, the stability of these gains is unclear. Appendix A.2.3 describes replacement of existing operations, but does not report insertion or deletion to correct an erroneous operation count; termination and length errors therefore remain relevant reliability questions. These gaps limit the evidence for meaningful error localization and improved geometric consistency, even though the reported system-level improvements are positive.

Questions for authors
  1. Please specify the evaluation counts and conversion/filtering rules, parameter-accuracy tolerances and denominators, treatment of sequence-length mismatches, the common validity criterion across output representations, and CD sampling, normalization, aggregation, and invalid-output handling. If standard implementations are used, please identify their versions and modifications. These details could materially affect confidence in Tables 1 and 2.

  2. Please document the adapted baselines' DINOv2 feature connections, freezing choices, versions, and training/tuning budgets. For Table 3, are the recovery variants identical in architecture and training settings apart from recovery direction and state supervision? How are differences in the image encoder, supervision, and inference computation controlled in the direct-generation comparison? These answers would affect the assessment of both comparative performance and the contribution of CCPR.

  3. Please define the exact criterion used to construct operation-error labels. Are existing measurements available for incorrect-to-correct and correct-to-incorrect changes after redrawing, termination and operation-count failures, geometric-relation residuals, solver fallback rates, and variability in GCR gains? If these have not been measured, please clarify which mechanism claims can be supported by the current evidence.

5 comments

Thank you for your insightful comments. We answer your concerns and questions in the following.

Q1-1. Please specify the evaluation counts and conversion/filtering rules. If standard implementations are used, please identify their versions and modifications.

A: We use the original DeepCAD evaluation implementation without modifying its metric scripts. The DeepCAD and Fusion 360 test splits contain 7,881 and 1,346 CAD models, respectively. All test models were successfully converted for evaluation, with no exclusions during data preparation. These evaluation details are provided in Appendix A.2.5.

Q1-2. Please specify the parameter-accuracy tolerances and denominators.

A: Command accuracy uses position-wise matching. Parameter accuracy considers valid parameters only at correctly predicted commands: continuous parameters require a quantized error <3, and categorical parameters require exact matching. The denominator of parameter accuracy is therefore the number of valid parameters at correctly predicted commands. Both accuracies are computed per sample and then averaged. Please see Appendix A.2.5.

Q1-3. Please specify the treatment of sequence-length mismatches.

A: If a prediction is shorter than the reference, the missing command positions are counted as mismatches. If it is longer, commands beyond the reference length are excluded from sequence accuracy. Concretely, we insert EOS at each prediction’s stopping position and compare predictions up to the reference EOS. Please see Appendix A.2.5.

Q1-4. Please specify the common validity criterion across output representations, as well as CD sampling, normalization, aggregation, and invalid-output handling.

A: The common validity criterion is successful conversion to a solid. CD sampling is 2,000 points per shape. Normalization is the same procedure across representations. Aggregation is the arithmetic mean of bidirectional squared nearest-neighbor distances over valid outputs. Invalid outputs (solid-construction or point-sampling failures) count toward IR and are excluded from CD; the IR denominator is all evaluated test models. Please see Appendix A.2.5.

Q2-1. Please document the adapted baselines’ DINOv2 feature connections, freezing choices, and versions.

A: For Drawing2CAD, CADrawer, and BrepDiff, which have no native raster or image interface, we use a frozen DINOv2 ViT-S/14 with registers. We replace only the input module and feed the features into the original downstream networks. For BrepDiff, the features are injected into its original denoising network as a conditioning signal.

Methods that natively accept images (GenCAD, Img2CAD, Img2CADSeq, CADRILLE, and CADReasoner) keep their original encoders and do not use DINOv2. HiFi-BRep already conditions on images and wireframe sketches via a pretrained DINOv2 model; we follow this native interface and retrain it on our DeepCAD raster sketches. Please see Appendix A.2.2.

Q2-2. Please document the training/tuning budgets.

A: Each compared method retains its own original training and tuning settings, which are recorded in Appendix A.2.4 and Table 5. Multi-stage methods retain their stage-wise schedules; we did not impose a shared epoch or search budget. Epochs, steps, and batch sizes are method-specific.

Q2-3. Free2CAD also receives ordered stroke information unavailable to raster-only methods. Please make this difference in input conditions explicit.

A: We have reported the difference in Appendix A.2.2 in the original submission, and added further discussions.

Q2-4. For Table 3, are the recovery variants identical in architecture and training settings apart from recovery direction and state supervision?

A: Yes. The three state-recovery variants in Table 3 use the same architecture, image encoder, and training settings. They differ only in recovery direction or intermediate-state supervision, and have identical parameter counts and computation (49.22M parameters; 33.61 GFLOPs).

Q2-5. How are differences in the image encoder, supervision, and inference computation controlled in the direct-generation comparison?

A: The image encoder is unified by using the same frozen DINOv2. Supervision differs in that the direct-generation variants have no intermediate-state supervision, while CCPR uses backward recovery with state supervision. Inference computation is not equalized; we report parameters, GFLOPs, and median latency.

Compared with CCPR, the encoding-stage parameters in direct generation are fully frozen, so we introduce two trainable Transformers for this comparison. We also report a direct-decoding experiment without these Transformers:

VariantACC_cmd ↑ACC_param ↑IR ↓CD ↓ParametersGFLOPsMedian latency
Direct decoding79.2%75.8%27.29%26.241.20M14.6213.87 ms
Direct generation80.2%76.6%23.21%23.155.54M16.4714.16 ms
CCPR81.3%81.7%4.81%18.749.22M33.61130.42 ms

CCPR improves reconstruction, particularly geometric validity, at higher inference cost. Please see Appendix A.3.3 and Table 6.

Q3-1. Please define the exact criterion used to construct operation-error labels.

A: An entire sketch–extrude operation is labeled erroneous if any command mismatches the reference, or if any valid parameter violates the evaluation criterion (continuous quantized error <3, categorical exact match). Please see Appendix A.2.3 for the operation-error label, and Appendix A.2.5 for the quantized-error tolerance and exact-match rule.

Q3-2. Are existing measurements available for error-detection precision and recall?

A: Yes. Error-detection precision is 98.89% and recall is 71.53%. Please see Appendix A.3.7 and Table 10.

Q3-3. Are existing measurements available for incorrect-to-correct and correct-to-incorrect changes after redrawing?

A: Yes. After redrawing, 41.42% of initially incorrect operations become correct, and 0.65% of initially correct operations become incorrect:

MeasureResult
Incorrect → correct after redrawing41.42%
Correct → incorrect after redrawing0.65%

Please see Appendix A.3.7 and Table 10.

Q3-4. Are existing measurements available for selection comparisons using the same redrawing model and initial predictions?

A: Yes. All-operation and selective redrawing use the same redrawing model and the same initial predictions; only operation selection differs.

Q3-5. Are existing measurements available for termination and operation-count failures? Can EAR insert or delete operations to correct an erroneous operation count?

A: Yes. Operation-count agreement with the reference is 84.11% for CCPR and 86.02% after EAR. EAR can terminate early and truncate subsequent operations, but cannot insert operations or independently delete an intermediate operation. Please see Appendix A.3.7.

Q3-6. Are existing measurements available for geometric-relation prediction quality, residuals, and constraint satisfaction?

A: Yes. Relation prediction on actual EAR outputs achieves 97.61% accuracy and 97.10% macro-F1. Precision and recall are 92.53% / 99.92% for parallel relations and 94.33% / 99.98% for perpendicular relations.

Angular residualBefore GCR: meanAfter GCR: mean
Parallel0.1490°0.1063°
Perpendicular0.1794°0.1303°
Overall0.1640°0.1182°

GCR reduces the overall mean angular residual by 27.9%, which is our direct measure of improved constraint satisfaction. Please see Appendix A.3.7 and Table 11.

Q3-7. Are existing measurements available for solver fallback rates and variability in GCR gains?

A: Yes. Of 5,681 solver calls, 54 did not converge and reverted to the initial parameters, giving a 0.95% fallback rate. Across three independently trained random seeds, we obtain:

MetricWithout GCRWith GCR
ACC_param (%) ↑82.2 ± 0.031982.3 ± 0.0342
CD ↓14.3 ± 0.007013.7 ± 0.0067
IR (%) ↓3.96 ± 0.08873.88 ± 0.0730

Please see Appendix A.3.7 and Table 12.

Reviewer_MNTxOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CADFlow, a sketch-to-CAD framework that recovers intermediate construction-state features and decodes them into parametric CAD sequences. It further introduces error-aware redrawing and geometric constraint refinement to improve reconstruction accuracy and validity. Experiments on DeepCAD, Fusion 360, and hand-drawn sketches demonstrate its effectiveness.

Critical strengths
  1. CADFlow achieves substantial improvements in geometric validity and competitive reconstruction accuracy. The cross-dataset and hand-drawn evaluations further support its effectiveness beyond the training distribution.
  2. CAD Construction Process Recovery (CCPR) introduces intermediate construction-state features to supervise and guide autoregressive CAD generation. The ablations demonstrate the benefits of backward recovery and intermediate-state supervision, particularly for geometric validity.
Critical weaknesses
  1. CADFlow is supervised against a single reference modeling history, although multiple modeling strategies can produce the same geometry. The method does not explicitly account for modeling-sequence quality or differing design requirements.
  2. The method supports only sketch-and-extrude operations, limiting its modeling expressiveness.
Questions for authors
  1. Could the framework be extended to generate different modeling sequences for the same geometry under different design requirements ? A conceptual discussion would suffice.
  2. How are the eight sampled views per model used during training, and are test viewpoints drawn from the same distribution? Is there evidence of generalization to other viewing conditions?
1 comment

Thank you for your insightful comments. We answer your concerns and questions in the following.

W1/Q1. The method does not explicitly account for modeling-sequence quality or differing design requirements. Could the framework be extended to generate different modeling sequences for the same geometry under different design requirements? A conceptual discussion would suffice.

A: Yes, a potential extension could be to add an extra text prompt that specifies the design requirement together with the input sketch, such as fewer operations, or a preferred feature order. An MLLM encoder (e.g., QWen) can embed the text into an implicit requirement code, which further attends to the features in the CCPR layers. At the data side, we can further process data samples with the corresponding requirement texts. With the mentioned extension, the method could potentially support the requirement specification.

W2. The method supports only sketch-and-extrude operations, limiting its modeling expressiveness.

A: We acknowledge this limitation. CADFlow currently supports only sketch-and-extrude operations. However, this representation can already express many complex models through successive operations, including holes, grooves, local cuts, and connected structures, as illustrated in Figures 4–5 and the Fusion360 reconstruction dataset. Future work will explore richer operation representations and corresponding construction-trajectory data to support revolutions, sweeps, and lofts.

Q2-1. How are the eight sampled views per model used during training, and are test viewpoints drawn from the same distribution?

A: During training, we randomly sample eight views per CAD model with elevations in [−45°, 45°]. Each view and its construction trajectory form an independent training sample; the model receives only one sketch at a time. The original test views follow the same sampling distribution.

Q2-2. Is there evidence of generalization to other viewing conditions?

A: We evaluate two additional views per test model outside the training elevation range: one from [−60°, −45°) and one from (45°, 60°].

Test viewpointsACC_cmd ↑ACC_param ↑CD ↓IR ↓
Original test distribution82.3%82.4%13.73.88%
Outside training elevation range75.7%78.5%20.93.87%

The main contours remain visible, so the model can still recover the basic object shape, and IR is nearly unchanged (3.88% → 3.87%). However, the projection has changed, and such elevations were not seen during training, so command accuracy, parameter accuracy, and CD drop, while a reasonable level of accuracy is still retained. Please see Appendix A.3.5 and Table 8.

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

This paper proposes CADFlow, a framework for generating executable parametric CAD models from raster sketch images. Its core component, CAD Construction Process Recovery (CCPR), recursively recovers intermediate construction-state features in the backward direction and then decodes adjacent-state transitions into CAD operations. An Error-Aware Redrawing (EAR) module further improves reconstruction quality through selective operation correction and geometric constraint refinement. Experiments on DeepCAD and Fusion 360 demonstrate improvements in reconstruction accuracy, geometric validity, and cross-dataset generalization. The paper also includes component ablations, hand-drawn sketch evaluations, and computational cost analysis.

Critical strengths
  1. Effective and well-motivated method design. The backward construction-state recovery and forward CAD decoding provide a meaningful approach to handling structural ambiguity in raster sketches. The combination with error-aware local redrawing is coherent and contributes to generating more reliable CAD sequences.
  2. Convincing experimental results. CADFlow consistently outperforms the compared methods across DeepCAD and Fusion 360 in reconstruction accuracy and geometric validity. The ablation studies provide useful evidence for the effectiveness of the proposed components, while the hand-drawn sketch evaluation further demonstrates the method's practical potential.
Critical weaknesses
  1. Limited geometric constraint coverage. Although CADFlow supports lines, arcs, and circles, its geometric constraint refinement only considers parallel and perpendicular relationships between line segments. Other common constraints, such as tangency and concentricity, are not considered.
  2. Comparability of adapted baselines. As described in Appendix A.2.2, several baselines are adapted to raster sketches using a shared DINOv2 encoder, while Free2CAD retains access to stroke-layer information. Although these adaptations are reasonable and clearly documented, the differences in input settings may affect the interpretation of the performance comparisons. A brief discussion of this aspect would make the evaluation more transparent.
Questions for authors

Here are some questions I want to discuss with the author(s):

  1. The hand-drawn sketches in the evaluation were produced by participants with CAD modeling experience. What about sketches drawn by novice users, which may be less precise or more ambiguous? How well would CADFlow handle such inputs?
  2. The full framework improves reconstruction quality but introduces additional inference time. How would this trade-off affect its use in interactive CAD modeling?
  3. I also curious about how CCPR handles the ambiguity of construction histories. A final CAD model may result from different valid construction sequences. If this ambiguity affects the backward recovery of intermediate construction states in CCPR? And how?
2 comments

Thank you for your insightful comments. We answer your concerns and questions in the following.

W1: Limited geometric constraint coverage. Although CADFlow supports lines, arcs, and circles, its geometric constraint refinement only considers parallel and perpendicular relationships between line segments. Other common constraints, such as tangency and concentricity, are not considered.

A: Our geometric constraint coverage (GCR) is not inherently restricted to parallel and perpendicular constraints. GCR can be extended to additional constraint types by incorporating the corresponding training annotations and optimization objectives.

To demonstrate this extensibility, we have extended GCR to support tangency and concentricity. The performance of the correction is shown below:

RelationBefore refinementAfter refinement
Tangency residual ↓0.6740.225
Center distance of concentric pairs ↓0.2840.048

Additionally, compared with the original GCR, the extended version also reduces CD from 13.7 to 13.5 and IR from 3.88% to 3.59%. These results show improved satisfaction of the added constraints, together with a modest improvement in reconstruction quality and validity. Please see Appendix A.3.6 and Table 9.

W2: Comparability of adapted baselines. As described in Appendix A.2.2, several baselines are adapted to raster sketches using a shared DINOv2 encoder, while Free2CAD retains access to stroke-layer information. Although these adaptations are reasonable and clearly documented, the differences in input settings may affect the interpretation of the performance comparisons. A brief discussion of this aspect would make the evaluation more transparent.

A: Thank you for this suggestion and have added a brief discussion in Appendix A.2.2 and in the following.

Methods that natively accept single-view images (GenCAD, Img2CAD and Img2CADSeq) are used with their original encoders and interfaces; this adaptation does not change their input setting. HiFi-BRep already conditions on images and wireframes via a pretrained DINOv2 model; we retain this native interface and retrain it on our DeepCAD raster sketches.

We then discuss the methods whose input settings were changed.

  • CADRILLE and CADReasoner originally accept multi-view images, but we replace them with the same single-view raster sketches and retrain them. The missing additional views might limit their view-aggregation process.
  • Drawing2CAD originally accepts SVG vector drawings, but we replace them with raster sketches. The missing exact 2D command coordinates might complicate its command encoding and parameter regression. After this adaptation, it remains the strongest CAD-sequence baseline (see Tab. 1) and its command accuracy drops slightly from 80.7% to 80.1%.
  • CADrawer originally accepts an ordered 3D sketch and a stroke/loop graph, but we replace them with a raster sketch. The missing stroke identities might hinder its primitive selection and parameter regression.
  • BrepDiff is originally an unconditional generator. We inject DINOv2 sketch features into its original denoising network to change it to sketch-conditioned generation. The extra semantic cues in the DINO feature could help guide the sampling process.
  • Free2CAD accesses the occluded sketches, but in our setting, these sketches are not drawn. The missing information might challenge their grouping and parameter regression processes.

Note that the extra DINOv2 feature provides semantic cues that might be helpful for these methods. The adaptation does not reverse the empirical conclusion: CADFlow outperforms both the unmodified raster-native methods and the adapted baselines.

Q1: The hand-drawn sketches in the evaluation were produced by participants with CAD modeling experience. What about sketches drawn by novice users, which may be less precise or more ambiguous? How well would CADFlow handle such inputs?

A: We have invited 10 novice users to draw their sketches of the same 200 CAD models used in the original hand-drawn study. Comparisons are shown below:

Sketch creatorsCD ↓IR ↓
CAD-experienced participants16.25.9%
Novice participants18.56.5%

CD rises from 16.2 to 18.5 and IR from 5.9% to 6.5%, so reconstruction quality is slightly worse on the novice sketches. Even so, 93.5% of the resulting CAD shapes remain valid, compared with 94.1% from CAD-experienced participants. This shows that CADFlow can handle less precise and more ambiguous novice sketches, however, we did not conduct the stress test to examine the feasible magnitude of the sketch roughness.

Quantitative and qualitative results are provided in Appendix A.3.4, Table 7, and Figure 11.

Q2: The full framework improves reconstruction quality but introduces additional inference time. How would this trade-off affect its use in interactive CAD modeling?

A: As reported in Tab. 5 in the original submission (it is Tab. 6 in the revision), the runtime of the full framework is 173.15 ms, with a breakdown of 130.42 ms for CCPR and 42.73 ms for EAR, making the whole method an interactive speed.

Q3. I also curious about how CCPR handles the ambiguity of construction histories. A final CAD model may result from different valid construction sequences. If this ambiguity affects the backward recovery of intermediate construction states in CCPR? And how?

A: In the training data, we indeed have different modeling sequences which lead to the same final CAD model. We treat each data sample equally, and do not explicitly handle the ambiguity. As a result, the model learns the most common modeling sequence of the same CAD model. CCPR is a deterministic process, thus is not affected by the intermediate construction states and cannot choose which construction sequence to generate.

Comments on the paper

3 comments

We sincerely thank all reviewers for their constructive and detailed feedback. We have carefully addressed every concern through clarification and additional experiments in our point-by-point responses below. We have revised the paper by incorporating all the new results, analyses, and revisions, and updated the PDF for your reference. All the edits are marked with blue text. We greatly appreciate the reviewers’ time and insightful comments, which have helped us substantially improve our work.

Dear Reviewers, the authors have posted their comments in response to the reviews. Please make an effort to discuss your review and the authors' response. - regards AC

AuthorsOpenReview ↗

Summary of Rebuttal

Dear Area Chair,

We sincerely thank you for overseeing the review process and all three reviewers for their thoughtful and constructive feedback. Throughout the rebuttal, we have carefully addressed the reviewers' concerns through detailed clarifications, additional experiments, and extended evaluations. These efforts have further strengthened the technical contributions and empirical validation of our work. We believe the revised manuscript provides a more comprehensive and convincing demonstration of the proposed framework's effectiveness and generalizability.

We particularly appreciate the reviewer who actively engaged in the discussion, acknowledged that their main concerns had been addressed, and raised their score from 2 to 3. The overall ratings have thus improved from 2/3/3 to 3/3/3. We sincerely appreciate the reviewers' valuable suggestions and the Area Chair's time and consideration. We hope that our detailed responses and the strengthened manuscript will be taken into account in the final assessment.

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.