acceptodds
Under review as a conference paper at ICLR 2027

DeCo: Efficient Decouple-to-Couple Learning for Multi-Task Visual Grounding

Abstract

Multi-task visual grounding requires models to jointly understand linguistic semantics and perform accurate visual localization and segmentation. Despite the success of multimodal large language models, effectively adapting them to multiple grounding objectives remains challenging. Existing methods commonly enforce task cooperation through shared representations, while overlooking the intrinsic conflict between task-oriented feature interests. In this paper, we introduce **DeCo**, an efficient **De**couple-to-**Co**uple learning framework that resolves this dilemma through a two-stage paradigm: task-specific representation decoupling followed by complementary prior coupling. Specifically, we first propose Task-aware Semantic Decoupling (TSD) to route shared visual cues into individual features under salient word-level guidance, alleviating representation interference between localization and segmentation. Furthermore, we observe that segmentation naturally provides informative localization priors due to dense supervision. Based on this insight, we introduce Hybrid Prior Coupling (HPC), which integrates sentence-level semantic prior with mask-derived spatial prior for enhanced grounding. Built upon a frozen multimodal encoder, DeCo requires lightweight trainable parameters while achieving strong generalization across multiple grounding objectives. Extensive experiments on RefCOCO/+, G-Ref, ReferIt, Flickr, DIOR-RSVG, SARVG1.0, RRSIS-D, RIS-LAD, and RefDIOR demonstrate that DeCo achieves state-of-the-art performance on both natural and remote sensing benchmarks. The code and model will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-Task Visual Grounding, Parameter-Efficient Fine-Tuning, Decouple-to-Couple Learning, Task-aware Semantic Decoupling, Hybrid Prior Coupling

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

DeCo jointly learns referring-expression localization and segmentation. It separates task-specific features from a shared vision–language encoder, then couples the branches using text features and a mask-derived spatial prior. A localization decoder predicts boxes and an auxiliary mask, while a UNet predicts the final segmentation mask. The training objective combines task losses with a mask-to-box loss. Experiments evaluate both tasks under dataset-specific and mixed training, including remote-sensing data and component comparisons.

Critical strengths
  1. The model performs well on both localization and segmentation, including against methods built on the same BEiT3 backbone family.

  2. The component comparisons show that decoupling and prior coupling each add value over naive joint training.

Critical weaknesses
  1. The claimed geometric feedback relies on a box extracted from a predicted binary mask. If that mask is hard-thresholded during training, the conversion blocks this gradient path and makes the mask-to-box loss ineffective for training the mask predictor; the paper gives no alternative backward rule.
Questions for authors
  1. What tensor is passed to mask-to-box during training: probabilities or a thresholded mask? Please show the backward rule and clarify whether the localization and mask-to-box losses update the UNet through this conversion.

  2. If the box is detached only before entering the localization decoder, while the separate mask-to-box loss is left unchanged, does segmentation accuracy change? This isolates feedback from localization rather than the benefit of using the mask as a forward prior.

  3. Do the DoM and full-coupling comparisons use the same auxiliary instance-mask supervision and loss weights? If not, can you separate the gain from extra supervision from the gain due to coupling?

  4. Do mask-derived boxes become accurate earlier than localization predictions during training? Curves for these two outputs would directly check the stated premise for the spatial prior.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors of DeCo

We sincerely thank you for the time and effort devoted to our paper and for your careful and constructive comments! We hope our point-by-point responses below can address your concerns, and we look forward to any further discussion.

Q1. What tensor is passed to mask-to-box during training: probabilities or a thresholded mask? Please show the backward rule and clarify whether the localization and mask-to-box losses update the UNet through this conversion.

A: Thanks for pointing out this issue. During training, we pass a thresholded binary mask rather than probabilities to the mask-to-box operation. We acknowledge that the original manuscript did not clearly distinguish this forward spatial prior, and have revised the relevant descriptions in Figure 1, Introduction, Sections 3.1, 3.3, and 3.4 accordingly.

To further investigate this issue, we conduct additional experiments with five M2B configurations (A–E), as reported in Section 4.3 (Effects of Mask-to-Box Variants), Table 9, and Figure 3(d). Specifically, we introduce a straight-through estimator (STE) based differentiable alternative that preserves hard box extraction in the forward pass while approximating backward gradients using probability-weighted spatial moments, as detailed in Appendix F. As shown in the following table, the original hard M2B (A) achieves the best performance (76.6%), while removing Lm2b\mathcal{L}_{m2b} yields comparable results (76.5%). The differentiable alternatives (C–E) achieve lower scores (75.0%–76.3%) and exhibit unstable convergence in Figure 3(d).

ModelsM2BLm2b\mathcal{L}_{m2b}Lm2b\mathcal{L}_{m2b} → RESLdet\mathcal{L}_{det} → RES[email protected][email protected]:.95oIoUmIoUAvg.
AHard✓××87.071.773.074.576.6
BHard×××86.771.673.274.576.5
CSTE✓✓✓86.170.371.472.275.0
DSTE✓✓×86.470.971.072.175.1
ESTE××✓86.871.273.174.076.3

These findings support retaining M2B as a forward geometric prior that transfers spatial knowledge from RES to REC, facilitating more accurate localization. We have clarified the use of Lm2b\mathcal{L}_{m2b} as an optional objective in Section 3.4.

Q2. If the box is detached only before entering the localization decoder, while the separate mask-to-box loss is left unchanged, does segmentation accuracy change? This isolates feedback from localization rather than the benefit of using the mask as a forward prior.

A: Thanks for raising this concern. We conduct the requested isolation experiment using STE-based M2B. Specifically, both configurations retain Lm2b\mathcal{L}_{m2b}, while the box is detached only before entering the localization decoder in configuration D.

Thus, D removes the gradient from Ldet\mathcal{L}_{det} to RES through M2B while preserving the gradient from the mask-to-box loss. As shown below, detaching the localization input results in only marginal changes in segmentation performance: oIoU decreases from 71.4% to 71.0%, while mIoU decreases from 72.2% to 72.1%. These results suggest that localization-to-segmentation gradient feedback provides limited additional benefits.

Configuration[email protected][email protected]:.95oIoUmIoUAvg.
C: Connected query86.170.371.472.275.0
D: Detached query86.470.971.072.175.1
AuthorsOpenReview ↗

Official Comment by Authors of DeCo

Q3. Do the DoM and full-coupling comparisons use the same auxiliary instance-mask supervision and loss weights? If not, can you separate the gain from extra supervision from the gain due to coupling?

A: Thanks for raising this concern. Yes, all experiments use the same auxiliary instance-mask supervision and loss weights, expert for LoM and SoM, as listed in the following table.

ModelsLdet\mathcal{L}_{det}Lins\mathcal{L}_{ins}Lseg\mathcal{L}_{seg}Lm2b\mathcal{L}_{m2b}
LoM✓✓××
SoM××✓✓
DoM✓✓✓✓
DoM w/ TSD✓✓✓✓
DoM w/ HPC✓✓✓✓
DeCo✓✓✓✓

We have added the corresponding description in the revised Section 4.3 (Effects of Core Components).

Q4. Do mask-derived boxes become accurate earlier than localization predictions during training? Curves for these two outputs would directly check the stated premise for the spatial prior.

A: Thanks for highlighting this suggestion. We illustrate additional convergence curves, as shown in the revised Section 4.3 (Effects of Core Components) and Figure 3(a)-(c).

Specifically, Figure 3(a) compares the validation performance of different model variants, showing that DeCo achieves consistently strong performance across REC and RES, with TSD and HPC providing complementary improvements over DoM. Figure 3(b) directly compares the localization accuracy of REC-predicted boxes from LoM and mask-derived boxes from SoM throughout training. The results show that mask-derived boxes provide more accurate localization cues during early training, although LoM gradually catches up as training progresses. Figure 3(c) further illustrates the corresponding loss convergence of the two branches.

These observations support our motivation for incorporating dense segmentation cues as spatial priors into the localization query through HPC, facilitating faster convergence and more accurate localization.

Thanks for clarifying that M2B works as a forward prior. The gradient ablations and added training curves address my main concerns. I’m satisfied with the response and will keep my score of 3 (Weak Accept).

AuthorsOpenReview ↗

Official Comment by Authors of DeCo

We sincerely thank Reviewer tAKu for the positive feedback, and we are glad that our additional experiments and clarifications have addressed your primary concern. Once again, we sincerely appreciate your valuable time and effort, responsible and thorough review, and constructive suggestions.

Reviewer_57oaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DeCo, a parameter-efficient framework for jointly performing referring expression comprehension and segmentation. Specifically, it first learns task-specific representations through Task-aware Semantic Decoupling and then transfers semantic and mask-derived spatial priors through Hybrid Prior Coupling. Experiments on natural-image and remote-sensing benchmarks report strong performance across both tasks.

Critical strengths
  1. The paper identifies a meaningful tension between task-specific representation learning and cross-task knowledge sharing. The proposed decouple-then-couple formulation is intuitive and reasonably simple.
  2. The author provides extensive experimental results. covering multiple natural-image and remote-sensing datasets under both standard and mixed-training settings.
  3. The paper is clear and easy to follow.
Critical weaknesses
  1. The central claim that shared representations cause task interference is not directly demonstrated. The evidence is mainly the performance difference between localization- and segmentation-oriented heads. Measurements of gradient conflict, representation similarity, or task affinity would provide stronger support.
  2. The differentiability of HPC is unclear. The paper describes converting a predicted binary mask into a box using global extrema. If hard binarization and min/max operations are used, it is unclear how the box-related losses can refine the segmentation branch as claimed.
  3. The current formulation is restricted to one referred object and one enclosing spatial prior, limiting applicability to multi-object or disconnected referents.
Questions for authors
  1. How is the predicted mask binarized before the mask-to-box operation?
  2. Is it possible to provide more evidence to support the central claim?
  3. Is it possible to extend the methods to multi-object task?
3 comments
AuthorsOpenReview ↗

Official Comment by Authors of DeCo

We sincerely thank you for the time and effort devoted to our paper and for your careful and constructive comments! We hope our point-by-point responses below can address your concerns, and we look forward to any further discussion.

Q1. How is the predicted mask binarized before the mask-to-box operation?

A: Thanks for pointing out this issue. The predicted mask logits are first converted into a probability map using a sigmoid function and then binarized with a threshold of 0.5, i.e., M=I(σ(S)≥0.5)\mathbf{M}=\mathbb{I}(\sigma(\mathbf{S})\geq 0.5).

We clarify that the original mask-to-box (M2B) operation applies hard thresholding and discrete spatial extrema, making it non-differentiable. We acknowledge that the original manuscript did not clearly distinguish this forward spatial prior, and have revised the relevant descriptions in Figure 1, Introduction, Sections 3.1, 3.3, and 3.4 accordingly.

To further investigate this issue, we conduct additional experiments with five M2B configurations (A–E), as reported in Section 4.3 (Effects of Mask-to-Box Variants), Table 9, and Figure 3(d). Specifically, we introduce a straight-through estimator (STE) based differentiable alternative that preserves hard box extraction in the forward pass while approximating backward gradients using probability-weighted spatial moments, as detailed in Appendix F. As shown in the following table, the original hard M2B (A) achieves the best performance (76.6%), while removing Lm2b\mathcal{L}_{m2b} yields comparable results (76.5%). The differentiable alternatives (C–E) achieve lower scores (75.0%–76.3%) and exhibit unstable convergence in Figure 3(d).

ModelsM2BLm2b\mathcal{L}_{m2b}Lm2b\mathcal{L}_{m2b} → RESLdet\mathcal{L}_{det} → RES[email protected][email protected]:.95oIoUmIoUAvg.
AHard✓××87.071.773.074.576.6
BHard×××86.771.673.274.576.5
CSTE✓✓✓86.170.371.472.275.0
DSTE✓✓×86.470.971.072.175.1
ESTE××✓86.871.273.174.076.3

These findings support retaining M2B as a forward geometric prior that transfers spatial knowledge from RES to REC, facilitating more accurate localization. We have clarified the use of Lm2b\mathcal{L}_{m2b} as an optional objective in Section 3.4.

Q2. Is it possible to provide more evidence to support the central claim?

A: Thanks for raising this suggestion. To provide more evidence of task interference between REC and RES, we conduct a task-gradient analysis on G-Ref (test-u). Specifically, for each test sample, we compute the cosine similarity between the gradients of the REC and RES losses with respect to the shared third-level feature V3\mathbf{V}_3.

As shown in the revised Section 4.3 (Analysis of Representation Interference) and Figure 4, the violin plot reveals that DoM exhibits gradient similarities concentrated around zero, whereas TSD and HPC shift the distributions toward positive values, with DeCo achieving the most favorable alignment. Consistently, the gradient conflict ratio (i.e., the proportion of samples with negative cosine similarity) decreases from 48.3% in DoM to 12.4% with TSD and 14.9% with HPC, further reaching 8.6% in DeCo, an 82.2% relative reduction over DoM.

ModelsConflict Ratio (%)
DoM48.3
DoM w/TSD12.4
DoM w/HPC14.9
DeCo8.6

Additionally, we provide Grad-CAM visualizations and analysis in Appendix D (Visualization of Task-specific Feature Preferences and Decoupling) to examine task-specific feature preferences. LoM focuses on sparse discriminative regions, whereas SoM covers broader target extents. In contrast, shared representations in DoM exhibit dominant-task or mixed activation patterns, while TSD effectively separates these preferences into localization- and segmentation-oriented features. These results provide direct support for our central claim.

AuthorsOpenReview ↗

Official Comment by Authors of DeCo

Q3. Is it possible to extend the methods to multi-object task?

A: Thanks for raising this concern. We acknowledge that the current formulation is designed for single-target referring expressions and is less suitable for multiple or spatially disconnected objects. We have discussed this limitation and potential extensions in Appendix E (Limitation and Future Works). Specifically, extending DeCo with multiple object queries and instance-aware mask decomposition would enable independent localization and spatial-prior extraction for each referred instance, facilitating multi-object visual grounding.

AuthorsOpenReview ↗

Official Comment by Authors of DeCo

Dear Reviewer 57oa,

Thank you again for your time, effort, and constructive feedback.

We would like to kindly follow up on our rebuttal and revised manuscript, in which we have highlighted the changes for your convenience. We fully understand this is a busy time, and would greatly appreciate it if you could take a moment to review our responses when time permits.

We sincerely hope our clarifications address your concerns. Should any questions or uncertainties remain, we would be more than happy to provide further clarification.

Best regards,

The Authors

Reviewer_6wtbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DeCo, a parameter-efficient framework for multi-task visual grounding that jointly addresses referring expression comprehension (REC) and referring expression segmentation (RES). The authors argue that the conventional shared-representation paradigm does not adequately account for the different feature preferences of localization and segmentation. To this end, DeCo adopts a decouple-to-couple design consisting of two main components. Task-aware Semantic Decoupling (TSD) constructs task-specific representations through salient world-guided attention and lightweight experts, while Hybrid Prior Coupling (HPC) incorporates sentence-level semantic information and mask-derived spatial priors into the localization query. Experiments on natural image and remote sensing benchmarks demonstrate competitive performance against existing methods.

Critical strengths
  1. The decouple-to-couple model design provides a clear approach to balancing task-specific representation learning and cross-task information sharing.
  2. The proposed method achieves strong performance across multiple natural-image and remote sensing benchmarks, demonstrating its effectiveness in diverse visual grounding scenarios.
  3. The paper provides extensive comparisons and ablations, showing improvements from TSD and HPC while maintaining parameter-efficient backbone adaptation.
Critical weaknesses
  1. The motivation is insufficiently validated. The performance differences among LoM, SoM, and DoM are relatively small, and DoM slightly outperforms both task-oriented baselines. These results directly demonstrate the claimed representation interference between RES and RSC.
  2. The methodological novelty is somewhat limited. TSD mainly combines conventional cross-attention, feature modulation, and task-specific experts, while HPC shares similarities with existing mask-guided cross-task interaction methods. The distinction from prior works, particularly C3VG and InstanceVG, could be clearer.
  3. The implementation of the mask-to-box operation is unclear. Equations (7)-(8) derive bounding box coordinates from a predicted binary mask, but the paper does not explain how gradients propagate through this operation. This is particularly important for the claimed cross-task optimization and the mask-to-box loss.
  4. The authors argue that dense segmentation supervision provides reliable localization priors earlier in training, but no convergence analysis is presented to support this assumption.
Questions for authors
  1. How is the mask-to-box operation implemented to ensure gradient propagation from the localization and mask-to-box losses to the segmentation branch?
  2. Can the authors provide more reliable evidence of representation interference between REC and RES, such as task-gradient similarity or quantitative feature analysis, to better support the motivation for TSD?
  3. What distinguishes the proposed decouple-to-couple mechanism from existing cross-task interaction approaches, particularly C3VG and InstanceVG?
  4. The HPC design assumes that segmentation produces more accurate spatial priors during early training. Can the authors provide convergence curves comparing the localization accuracy of mask-derived boxes and REC predictions?
3 comments
AuthorsOpenReview ↗

Official Comment by Authors of DeCo

We sincerely thank you for the time and effort devoted to our paper and for your careful and constructive comments! We hope our point-by-point responses below can address your concerns, and we look forward to any further discussion.

Q1. How is the mask-to-box operation implemented to ensure gradient propagation from the localization and mask-to-box losses to the segmentation branch?

A: Thanks for pointing out this issue. We clarify that the original mask-to-box (M2B) operation applies hard thresholding and discrete spatial extrema, making it non-differentiable. We acknowledge that the original manuscript did not clearly distinguish this forward spatial prior, and have revised the relevant descriptions in Figure 1, Introduction, Sections 3.1, 3.3, and 3.4 accordingly.

To further investigate this issue, we conduct additional experiments with five M2B configurations (A–E), as reported in Section 4.3 (Effects of Mask-to-Box Variants), Table 9, and Figure 3(d). Specifically, we introduce a straight-through estimator (STE) based differentiable alternative that preserves hard box extraction in the forward pass while approximating backward gradients using probability-weighted spatial moments, as detailed in Appendix F. As shown in the following table, the original hard M2B (A) achieves the best performance (76.6%), while removing Lm2b\mathcal{L}_{m2b} yields comparable results (76.5%). The differentiable alternatives (C–E) achieve lower scores (75.0%–76.3%) and exhibit unstable convergence in Figure 3(d).

ModelsM2BLm2b\mathcal{L}_{m2b}Lm2b\mathcal{L}_{m2b} → RESLdet\mathcal{L}_{det} → RES[email protected][email protected]:.95oIoUmIoUAvg.
AHard✓××87.071.773.074.576.6
BHard×××86.771.673.274.576.5
CSTE✓✓✓86.170.371.472.275.0
DSTE✓✓×86.470.971.072.175.1
ESTE××✓86.871.273.174.076.3

These findings support retaining M2B as a forward geometric prior that transfers spatial knowledge from RES to REC, facilitating more accurate localization. We have clarified the use of Lm2b\mathcal{L}_{m2b} as an optional objective in Section 3.4.

Q2. Can the authors provide more reliable evidence of representation interference between REC and RES, such as task-gradient similarity or quantitative feature analysis, to better support the motivation for TSD?

A: Thanks for raising this suggestion. To provide quantitative evidence of representation interference between REC and RES, we conduct a task-gradient analysis on G-Ref (test-u). Specifically, for each test sample, we compute the cosine similarity between the gradients of the REC and RES losses with respect to the shared third-level feature V3\mathbf{V}_3.

As shown in the revised Section 4.3 (Analysis of Representation Interference) and Figure 4, the violin plot reveals that DoM exhibits gradient similarities concentrated around zero, whereas TSD and HPC shift the distributions toward positive values, with DeCo achieving the most favorable alignment. Consistently, the gradient conflict ratio (i.e., the proportion of samples with negative cosine similarity) decreases from 48.3% in DoM to 12.4% with TSD and 14.9% with HPC, further reaching 8.6% in DeCo, an 82.2% relative reduction over DoM.

ModelsConflict Ratio (%)
DoM48.3
DoM w/TSD12.4
DoM w/HPC14.9
DeCo8.6

Additionally, we provide Grad-CAM visualizations and analysis in Appendix D (Visualization of Task-specific Feature Preferences and Decoupling) to examine task-specific feature preferences. LoM focuses on sparse discriminative regions, whereas SoM covers broader target extents. In contrast, shared representations in DoM exhibit dominant-task or mixed activation patterns, while TSD effectively separates these preferences into localization- and segmentation-oriented features. These results provide direct evidence supporting our motivation for task-aware representation decoupling.

AuthorsOpenReview ↗

Official Comment by Authors of DeCo

Q3. What distinguishes the proposed decouple-to-couple mechanism from existing cross-task interaction approaches, particularly C3VG and InstanceVG?

A: Thanks for raising this concern. Although these methods share the goal of joint localization and segmentation, their motivations and interaction mechanisms differ fundamentally.

C3VG adopts a coarse-to-fine framework that refines preliminary predictions through mask-guided interaction and explicit bidirectional consistency constraints. In contrast, DeCo focuses on representation-level decoupling before cross-task coupling. Specifically, TSD constructs task-oriented features through salient word-guided attention and lightweight experts, while HPC incorporates sentence-level semantics and mask-derived spatial priors into the localization query without imposing bidirectional consistency constraints.

InstanceVG focuses on instance-aware joint learning for generalized visual grounding, employing unified instance queries with reference points to achieve consistent point, box, and mask predictions. In contrast, DeCo explicitly separates localization- and segmentation-oriented representations and subsequently transfers complementary geometric knowledge from RES to REC through HPC.

Q4. The HPC design assumes that segmentation produces more accurate spatial priors during early training. Can the authors provide convergence curves comparing the localization accuracy of mask-derived boxes and REC predictions?

A: Thanks for highlighting this suggestion. We illustrate additional convergence curves, as shown in the revised Section 4.3 (Effects of Core Components) and Figure 3(a)-(c).

Specifically, Figure 3(a) compares the validation performance of different model variants, showing that DeCo achieves consistently strong performance across REC and RES, with TSD and HPC providing complementary improvements over DoM. Figure 3(b) directly compares the localization accuracy of REC-predicted boxes from LoM and mask-derived boxes from SoM throughout training. The results show that mask-derived boxes provide more accurate localization cues during early training, although LoM gradually catches up as training progresses. Figure 3(c) further illustrates the corresponding loss convergence of the two branches.

These observations support our motivation for incorporating dense segmentation cues as spatial priors into the localization query through HPC, facilitating faster convergence and more accurate localization.

AuthorsOpenReview ↗

Official Comment by Authors of DeCo

Dear Reviewer 6wtb,

Thank you again for your time, effort, and constructive feedback.

We would like to kindly follow up on our rebuttal and revised manuscript, in which we have highlighted the changes for your convenience. We fully understand this is a busy time, and would greatly appreciate it if you could take a moment to review our responses when time permits.

We sincerely hope our clarifications address your concerns. Should any questions or uncertainties remain, we would be more than happy to provide further clarification.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Comment by Authors of DeCo

Dear Reviewers, Area Chairs, Senior Area Chairs, and Program Chairs,

We sincerely thank all reviewers for their time and thoughtful feedback. We are encouraged that our work has been recognized for its intuitive decouple-to-couple design, strong performance across diverse benchmarks, and extensive experimental evaluations.

The main concerns involve the validation of representation interference, the gradient propagation of Mask2Box, and the early localization advantage of mask-derived spatial priors. In response, we have strengthened our experimental and methodological analyses by adding quantitative task-gradient analysis (Section 4.3), reducing the conflict ratio from 48.27% in DoM to 8.64% in DeCo, along with complementary feature visualizations and quantitative analysis. We have also clarified the non-differentiability of Hard M2B (Introduction, Sections 3.1, 3.3, and 3.4), introduced an STE-based differentiable alternative (Section 3.4, Section 4.3, and Appendix F), and illustrated convergence analyses (Section 4.3) comparing mask-derived boxes with REC predictions.

We have provided detailed point-by-point responses and included the additional experimental results in the rebuttal PDF. The manuscript has also been carefully revised, with corresponding changes highlighted in red. We sincerely hope that our revisions can address the reviewers' concerns and greatly appreciate their constructive feedback, which has helped improve the quality of our paper.

Best regards,

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.