acceptodds
Under review as a conference paper at ICLR 2027

Vision Fluxer with Dynamic Global Receptive Field

Abstract

A token mixer should provide a global receptive field and input-dependent mixing at tractable cost: convolution, with its local, static kernel, offers neither, and self-attention offers both at a cost quadratic in the number of tokens. Frequency-domain convolution attains a global receptive field at , but its filter is fixed after training. Conditioning the filter pointwise on the input admits two constructions: the filter may be generated in frequency from the spectrum, or in space from the signal and then transformed. A linear generator cannot distinguish the two, and the choice between them has gone unexamined. Here we show that once the generator is nonlinear the two constructions are no longer interchangeable, and adopt the spatial one: Fluxer generates a full-resolution kernel as a pointwise projection of the input, so that the mixing amounts to a bilinear interaction between two nonlinear views of a single feature map. Three separations position Fluxer relative to static spectral filters, local convolution, and frequency-pointwise masks; our experiments bear most on the last: in the same backbone, a frequency-pointwise generator recovers only a small part of the gap between a static filter and Fluxer on ImageNet-1K. As a drop-in backbone, Vision Fluxer (ViF) achieves the most favorable accuracy–compute trade-off among representative single-stream hierarchical backbones on classification, detection, and segmentation, indicating that convolution, once equipped with a dynamic global kernel, constitutes a competitive foundation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
vision backbone, token mixer, dynamic convolution, global receptive field, Fourier transform, frequency-domain filtering, efficient architecture

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SU9kOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a visual backbone solution named Vision Fluxer to combine the idea of global filters in the frequency domain and adaptive/generated kernels. Although there are several methods working on adaptive/generated kernels, including AFFNet and CDFFormer, the proposed method emphasizes the importance of the coefficient of variation of the feature maps. The method is evaluated on widely used visual backbone benchmarks including ImageNet classification, COCO detection, and ADE20k segmentation.

Critical strengths
  • The idea is well-motivated and easy to follow.

  • The paper provides some interesting and in-depth analysis of the value of adaptive kernels for global filters.

Critical weaknesses
  • The paper highlights its performance (like Figure 1), and I believe superior performance is also key to supporting the value of the proposed method and claims. However, some important recent models are not compared/mentioned in the paper, and the proposed solution cannot outperform them. For example, the proposed method achieves similar performance on ImageNet to CAFormer and VSSD, which were published in 2024 and 2025, respectively. Interestingly, VSSD and the proposed method use similar training recipes and tricks like MESA. So it is still unknown whether the high performance comes from better architectural designs or better training strategies.

  • I think it is a good paper with a comprehensive experimental study and carefully designed analysis. However, I am a bit doubtful about the value of publishing a visual backbone paper nowadays with limited performance improvement and similar overall architectures.

Questions for authors

Please see my comments above.

After rebuttal: The authors add comparisons with more recent methods. However, I remain concerned that the improvements over existing methods are relatively marginal in terms of both technical value and empirical performance. Therefore, I will maintain my initial rating.

4 comments
AuthorsOpenReview ↗

Response to Reviewer SU9k

Thank you for recognizing the motivation and the analysis of adaptive global filters. We appreciate your suggestions on the comparisons and training recipe.

Comparisons with recent backbones

We provide comparisons with CAFormer, VSSD, and 2026 methods. Sources are CAFormer, VSSD, ICCV 2025, Random Wins All, CVPR 2026, and SPANetV2, TPAMI 2026. ViF results below are from our submission.

ImageNet-1K at 224 × 224. We include the S18/S36 variants of CAFormer and SPANetV2-hybrid. Entries are parameters (M) / FLOPs (G) / top-1 (%).

ModelParams (M)Compute (G)Top-1 (%)
CAFormer-S18264.183.6
CAFormer-S36398.084.5
VSSD-T285.083.8
VSSD-S508.184.6
VSSD-B8916.185.4
Random-Swin-T284.582.7
Random-Swin-S508.783.9
Random-Swin-B8815.484.4
SPANetV2-S18-hybrid274.283.9
SPANetV2-S36-hybrid418.184.7
ViF-T294.883.9
ViF-S518.884.8
ViF-B9916.985.4

CAFormer and SPANetV2-hybrid combine different token mixers, including attention, across stages. ViF uses Fluxer as the token mixer in all four stages, providing dynamic global convolution throughout the backbone. These comparisons assess complete backbones, whereas Fig. 5 isolates our kernel-generation design within a fixed backbone and training recipe.

COCO with Mask R-CNN, 1× schedule. Parameters/FLOPs include the detector.

ModelParams (M)FLOPs (G)Box APMask AP
VSSD-T5229847.042.8
Random-Swin-T4826746.041.9
ViF-T4928747.943.1
VSSD-S7738148.943.9
Random-Swin-S6935948.043.2
ViF-S7137549.544.3
VSSD-B10850649.944.8
Random-Swin-B10749649.144.6
ViF-B11753650.745.2

ADE20K with UPerNet. Parameters/FLOPs include the segmentor; SS/MS denote single-/multi-scale testing.

ModelParams (M)FLOPs (G)SS mIoU (%)
VSSD-T5995348.7
Random-Swin-T6094546.8
ViF-T6096349.2
VSSD-S81102150.3
Random-Swin-S81103848.9
ViF-S83105450.8

Token mixer and training recipe

Our contribution is the Fluxer token mixer and its analysis. We use an established four-stage hierarchy to evaluate its effectiveness as a general mixing mechanism. Developing such operators remains valuable because they provide reusable building blocks and design insights for visual representation learning beyond a particular backbone configuration.

Fluxer synthesizes a full-resolution convolution kernel from each image and applies it by FFT. It preserves convolutional weight sharing while making global interactions content-dependent. The nonlinear spatial construction permits cross-frequency interactions beyond frequency-pointwise kernel generation. We analyze this distinction and test the generator choices with the surrounding backbone and training recipe fixed.

We adopt MESA as an existing training technique. Without MESA, ViF-T/S/B reach 83.68/84.60/85.35%; with MESA, they reach 83.93/84.84/85.41%, giving gains of 0.25/0.24/0.06 points.

For Fig. 5, we retrain the generator variants with the same stem, downsampling, positional encodings, gated MLPs, widths, depths, and training recipe, including MESA. The static filter reaches 83.2%, frequency-pointwise generation 83.3%, shift-invariant generation 83.5%, and Fluxer 83.9%. The static-to-Fluxer gain is therefore measured with the surrounding architecture and recipe fixed.

Dear Reviewer SU9k,

Thank you again for your review. Even with self-attention firmly established, approaches such as Mamba and FNO have opened up new possibilities for model design by exploring different computational building blocks. This is also the motivation behind Fluxer: exploring global interaction through convolution, with a kernel generated directly from the current image features.

The paper develops this idea into a mixer that retains convolutional weight sharing and FFT evaluation, and studies the consequences of how its kernel is generated. We analyze what the construction allows and the restrictions it imposes, then examine it within a standard backbone across visual tasks. We hope this gives subsequent work a concrete alternative for global mixing, together with an understanding of its design that can guide further development.

We believe the value of this work lies in that operator and the knowledge needed to build on it, extending beyond a particular backbone configuration. We would be grateful if you could reconsider your rating in light of this contribution and the evidence in our response. We would also welcome any remaining concern you would like us to address.

Best regards,

The Authors

Thank you for the detailed feedback and additional comparisons. However, I remain concerned that the improvements over existing methods are relatively marginal in terms of both technical value and empirical performance. Therefore, I will maintain my initial rating.

Dear Reviewer SU9k,

Thank you for taking the time to consider our responses and the additional comparisons. We appreciate your careful review and your recognition of the experimental study and analysis.

We understand your remaining reservations about the significance of the contribution and would like to offer some further context. CAFormer, SPANetV2-hybrid, and VSSD combine attention with other mixers. ViF uses Fluxer as the token mixer in all four stages, with ViF-B reaching 85.4% top-1 accuracy on ImageNet-1K at 224 × 224. As reported in Section 5.3, Fig. 6 (left), and Table A8, the same ViF-T checkpoint trained at 224 × 224 is evaluated at the tested resolutions from 128 × 128 to 512 × 512 without fine-tuning or weight resampling. At 384 × 384, it achieves 82.97% top-1 accuracy, only 0.90 percentage points below its 83.87% at 224 × 224. We hope these properties, together with the analysis and controlled comparisons, offer useful additional context for evaluating Fluxer as a reusable global mixer.

We would also be grateful if you could clarify which specific experiments, comparisons, or aspects of the technical contribution remain a concern.

Best regards,

The Authors

Reviewer_YK9VOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper describes a model architecture for vision tasks. The central question is how tokens are mixed. The authors develop a comparatively efficient approach to performing global convolutions with input-dependent kernels. Specifically, full-resolution kernels are generated in the spatial domain, and an FFT-based construction performs the convolution efficiently. While dynamic convolutions, global Fourier-based convolutions, and input-dependent filters are all established ideas, the paper introduces a new point in the design space. The approach is evaluated through a reasonably convincing set of experiments, demonstrating competitive performance and modest improvements over strong baselines. However, the extent to which these improvements can be attributed specifically to the proposed kernel-generation mechanism or explained by the developed theory remains somewhat unclear.

Critical strengths

Given the ubiquity of attention mechanisms, with their quadratic complexity, in current model backbones, research like this is important. The proposed approach offers an interesting combination of global receptive fields, input-dependent mixing, and computational efficiency.

I find this to be an educational paper with considerable didactic value. I particularly like Fig. 2, which nicely communicates the central qualities this work aims to achieve. I further appreciate that the scope and limitations of this work are clearly communicated.

Also worth mentioning, of course, are the competitive results. The proposed design deserves to be taken seriously and is likely to inspire future work.

Critical weaknesses

The central motivation behind this work is not entirely transparent to me and sometimes seems hidden behind the prose. In the age of CNNs, two of the main reasons for using convolution were 1) (parameter) efficiency and 2) translation equivariance. It seems that this work primarily focuses on a variant of 1). Imposing a convolutional structure allows the FFT to be used as a computational shortcut. However, the input dependence of the kernel sacrifices property 2). Roughly speaking, two objects in an image are not guaranteed to produce the same response when they appear at different locations (Proposition 4). On the other hand, if two token pairs have the same relative displacement, they receive the same convolutional mixing weight, which represents a substantial restriction in flexibility compared to attention. So there is quite a price to pay for the proposed structure. I would encourage the authors to consider putting this central trade-off more front and center in the exposition.

Fig. 5 reveals that the difference between a static kernel and the full approach is 0.7 percentage points. This difference is smaller than the reported advantage over some older backbones (Table 1). The authors acknowledge that this comparison is not entirely clean because of various confounding factors. Nevertheless, this leaves some uncertainty about how much of the reported performance advantage can be attributed specifically to the proposed mechanism, and to what extent the theoretical differences explain the empirical improvements. I might have misunderstood something, but I would like to see a stronger comparison against a similarly parameterized dynamic global mixer that does not enforce convolutional structure.

Questions for authors

Please see my comments in the Weaknesses section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer YK9V

Thank you for your thoughtful review and your encouraging assessment of the analysis and Fig. 2.

Translation equivariance

Our aim is global, input-dependent mixing with the computational structure of convolution. Within each channel, token pairs with the same relative displacement share a mixing coefficient. The kernel changes across images but is shared across query locations. This restricts query-specific weighting and enables FFT evaluation in O(N log N) time per channel.

Our spatial kernel generator retains phase and positional information, at the cost of exact translation equivariance (Proposition 4). A translation-invariant generator can preserve equivariance; the global mixing core is circulant at each fixed input in both cases. In Fig. 5, the shift-invariant, phase-suppressing generator reaches 83.5% top-1 accuracy and Fluxer reaches 83.9%.

We additionally evaluated translations on all 50,000 ImageNet-1K validation images. Prediction consistency measures whether the predicted class remains unchanged. Under a 16-pixel cyclic translation, consistency is 96.4% for ViF-T, 94.4% for Swin-T, and 94.6% for ConvNeXt-T. Under a 16-pixel crop-window shift, which admits new content, the rates are 97.0%, 94.3%, and 95.7%. ViF-T accuracy changes from 83.87% to 83.54% under the cyclic shift, a 0.33-point decrease, versus 0.49 and 0.95 points for Swin-T and ConvNeXt-T. Its accuracy under the crop-window shift is 83.84%.

Effect of the dynamic kernel

The confounding-factor caveat concerns comparisons between independently developed backbones. In Fig. 5, we retrain the variants with the same stem, downsampling layers, positional encodings, gated MLPs, widths, depths, and training recipe. MESA is used in every variant. The static filter reaches 83.2%, the frequency-pointwise generator 83.3%, the shift-invariant generator 83.5%, and Fluxer 83.9%.

The 0.7-point gain therefore measures the generator change with the surrounding backbone and recipe fixed. The separately measured MESA gains are 0.25/0.24/0.06 points for ViF-T/S/B; MESA does not explain the static-to-dynamic gap because both variants use it.

Non-convolutional global mixers

MLLA, VSSD, Spatial-Mamba, and ViF share the same four-stage hierarchical backbone framework, with an overlapped convolutional stem, overlapped downsampling, and residual blocks combining a token mixer with an FFN. Their input-dependent global mixers use linear attention, non-causal state-space duality, structure-aware state-space mixing, and FFT convolution, respectively. We compare these backbones across classification, detection, and segmentation below, with model budgets included.

ViF values are from our submission. Baselines are from MLLA/MILA, Tables 3–4, VSSD, Tables 1 and 3–4, and Spatial-Mamba, Tables 1–3. VSSD denotes the ICCV 2025 version throughout.

ImageNet-1K at 224 × 224. Entries are parameters (M) / FLOPs (G) / top-1 (%).

BackboneTinySmallBase
MLLA/MILA25 / 4.2 / 83.543 / 7.3 / 84.496 / 16.2 / 85.3
VSSD28 / 5.0 / 83.850 / 8.1 / 84.689 / 16.1 / 85.4
Spatial-Mamba27 / 4.5 / 83.543 / 7.1 / 84.696 / 15.8 / 85.3
ViF29 / 4.8 / 83.951 / 8.8 / 84.899 / 16.9 / 85.4

COCO with Mask R-CNN, 1× schedule. Parameters/FLOPs include the detector.

BackboneParams (M)FLOPs (G)Box APMask AP
MLLA/MILA-T4425546.842.1
VSSD-T5229847.042.8
Spatial-Mamba-T4626147.642.9
ViF-T4928747.943.1
MLLA/MILA-S6331949.244.2
VSSD-S7738148.943.9
Spatial-Mamba-S6331549.244.0
ViF-S7137549.544.3
MLLA/MILA-B11550250.545.0
VSSD-B10850649.944.8
Spatial-Mamba-B11549450.445.1
ViF-B11753650.745.2

ADE20K with UPerNet. Parameters/FLOPs include the segmentor; SS/MS denote single-/multi-scale testing.

BackboneParams (M)FLOPs (G)mIoU SSmIoU MS
VSSD-T5995348.749.7
Spatial-Mamba-T5793648.649.4
ViF-T6096349.249.8
VSSD-S81102150.351.0
Spatial-Mamba-S7399250.651.4
ViF-S83105450.851.4

Dear Reviewer YK9V,

Thank you again for your thoughtful review and your encouraging comments on the analysis. Our reply explains the weight-sharing and equivariance trade-offs and reports both prediction consistency and accuracy under translations.

To examine non-convolutional global mixers, we compare ViF with MLLA, VSSD, and Spatial-Mamba, which use the same four-stage backbone framework and closely aligned training protocols. The comparisons span classification, detection, and segmentation. Within ViF, we also compare kernel generators with the surrounding backbone and training recipe held fixed.

Please let us know whether any aspect of the trade-off or the comparisons still needs clarification. We would appreciate it if you could revisit your rating based on the evidence in our response.

Best regards,

The Authors

I thank the authors for their detailed response. My concerns have been adequately addressed, and I maintain my original positive-leaning rating.

Thank you for your follow-up and for taking the time to consider our responses.

Reviewer_xpbGOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper designs a dynamical and global filter to process vision tokens using the fast Fourier transform. The results on classification, detection, and segmentation tasks are comparable with state-of-the-art methods.

Critical strengths

None

Critical weaknesses
  1. The operator is essentially an input-dependent kernel produced by pointwise spatial projection/nonlinearity and applied through FFT convolution. This is a sensible synthesis of dynamic filtering and spectral global mixing. However, it is very close to Dynamic Filter Networks [1], CondConv [2], Involution [3], GFNet [4], and similar works. The authors should describe their differences more thoroughly, and I am not sure whether the main performance gain comes from the architecture or just from modern techniques [5].

  2. The authors should compare their method with 2026 methods, the comparable methods in this paper are all earlier than 2025.

[1] Decoupled Dynamic Filter Networks. CVPR 2021. [2] CondConv: Conditionally Parameterized Convolutions for Efficient Inference. NIPS 2019. [3] Involution: Inverting the Inherence of Convolution for Visual Recognition. 2021. [4] GFNet: Global Filter Networks for Visual Recognition. TPAMI 2023. [5] A ConvNet for the 2020s. CVPR 2022.

Questions for authors

None

3 comments
AuthorsOpenReview ↗

Response to Reviewer xpbG

Thank you for your careful review and constructive feedback. We appreciate the time and effort you devoted to evaluating our work, and the concrete references and suggestions you provided. Your comments have helped us clarify the relationship to prior methods and the evidence for our design choices.

Relation to prior work and source of the gains

Kernel design

Fluxer generates a kernel covering the full feature grid from the current feature map. It applies this kernel by two-dimensional FFT convolution, allowing each output to aggregate all spatial locations in one mixing layer. The kernel is regenerated for each input and shared across spatial locations within an image.

The table separates kernel adaptivity from receptive field. Static kernels retain the same coefficients across inputs; dynamic kernels depend on the input. Receptive field denotes direct feature aggregation in a single layer.

MethodKernelConstructionReceptive field
ConvNeXtStatic, input-independentLearned depthwise kernelLocal
DDFDynamic, input-dependentSpatial and channel filter factorsLocal
CondConvDynamic, input-dependentInput-dependent mixture of static expertsLocal
InvolutionDynamic, input-dependentPosition-specific kernel generationLocal
GFNetStatic, input-independentLearned frequency weightsGlobal
AFFDynamic, input-dependentFrequency-pointwise generationGlobal
OrchidDynamic, input-dependentTranslation-invariant conditioningGlobal
FluxerDynamic, input-dependentSpatial pointwise nonlinear generationGlobal

DDF, CondConv, and Involution make local filtering input-dependent. Fluxer extends direct aggregation to the full feature map. GFNet already provides global aggregation, but its filter is fixed across inputs. In CondConv, the expert kernels are static, while their input-dependent combination makes the applied kernel dynamic.

The closer spectral relatives are AFF and Orchid. AFF generates its filter pointwise in frequency; Fluxer generates it pointwise in space before the transform. With a nonlinear generator, spatial generation can couple input frequencies. Orchid uses translation-invariant conditioning, whereas Fluxer retains spatial phase. Our simplified-core analysis examines these structural differences, and Fig. 5 compares the corresponding generator designs within the same backbone.

Architecture and training

ViF builds on an established four-stage hierarchical architecture. For the generator comparison in Fig. 5, we retrain the variants while holding the stem, downsampling layers, positional encodings, gated MLPs, widths, depths, and training recipe fixed. MESA is used in every variant. The ImageNet-1K results are:

Kernel generatorTop-1 (%)
Static spectral filter83.2
Frequency-pointwise, AFF-like83.3
Shift-invariant, Orchid-like83.5
Fluxer83.9

The 0.7-point static-to-Fluxer improvement is therefore measured with the surrounding architecture and training recipe fixed. The AFF-like and Orchid-like variants further compare the choice of generator within this shared backbone.

We also trained each ViF scale without MESA to measure its contribution separately:

ModelWithout MESAWith MESAGain
ViF-T83.6883.93+0.25
ViF-S84.6084.84+0.24
ViF-B85.3585.41+0.06
AuthorsOpenReview ↗

Response to Reviewer xpbG (continued)

Comparisons with recent methods

We provide comparisons with Random Wins All, CVPR 2026, SPANetV2, TPAMI 2026, CAFormer, and VSSD, ICCV 2025. Baseline results are quoted from these sources; ViF results are from our submission except for the additional Cascade experiments identified below.

ImageNet-1K at 224 × 224. We include the S18/S36 variants of CAFormer and SPANetV2-hybrid. Entries are parameters (M) / FLOPs (G) / top-1 (%).

ModelParams (M)Compute (G)Top-1 (%)
CAFormer-S18264.183.6
CAFormer-S36398.084.5
VSSD-T285.083.8
VSSD-S508.184.6
VSSD-B8916.185.4
Random-Swin-T284.582.7
Random-Swin-S508.783.9
Random-Swin-B8815.484.4
SPANetV2-S18-hybrid274.283.9
SPANetV2-S36-hybrid418.184.7
ViF-T294.883.9
ViF-S518.884.8
ViF-B9916.985.4

CAFormer and SPANetV2-hybrid combine different token mixers, including attention, across stages. ViF uses Fluxer as the token mixer in all four stages, providing dynamic global convolution throughout the backbone. These comparisons assess complete backbones, whereas Fig. 5 isolates our kernel-generation design within a fixed backbone and training recipe.

COCO with Mask R-CNN, 1× schedule. Parameters/FLOPs include the detector.

ModelParams (M)FLOPs (G)Box APMask AP
VSSD-T5229847.042.8
Random-Swin-T4826746.041.9
ViF-T4928747.943.1
VSSD-S7738148.943.9
Random-Swin-S6935948.043.2
ViF-S7137549.544.3
VSSD-B10850649.944.8
Random-Swin-B10749649.144.6
ViF-B11753650.745.2

COCO with Cascade Mask R-CNN, 3× schedule and multi-scale training. Parameters/FLOPs include the detector. Ego results are from Ego, CVPR 2026, Table 5. The ViF Cascade results are additional experiments not included in the original submission.

ModelParams (M)FLOPs (G)Box APMask AP
Ego-T8272850.744.0
ViF-T8776552.245.2
Ego-S9580352.145.1
ViF-S10985352.945.9
Ego-B11190953.345.8
ViF-B154101553.646.6

FLOPs are reported at 800 × 1280. Training recipes follow each method; Ego uses batch size 8 and ViF uses 16.

ViF improves box AP by 1.5/0.8/0.3 points and mask AP by 1.2/0.8/0.8 points at the T/S/B scales, with the model budgets listed above.

ADE20K with UPerNet. Parameters/FLOPs include the segmentor; SS denotes single-scale testing.

ModelParams (M)FLOPs (G)SS mIoU (%)
VSSD-T5995348.7
Random-Swin-T6094546.8
ViF-T6096349.2
VSSD-S81102150.3
Random-Swin-S81103848.9
ViF-S83105450.8

Dear Reviewer xpbG,

Thank you again for your careful review and the references you suggested. Our responses clarify how Fluxer differs from local dynamic filters, static global filters, and other dynamic spectral constructions.

Our evidence includes kernel-generator comparisons with the surrounding backbone and training recipe fixed, and comparisons with MLLA, VSSD, and Spatial-Mamba, which share the same four-stage backbone framework and closely aligned training protocols. We also provide comparisons with recent methods across classification, detection, and segmentation, including additional Cascade experiments not included in the original submission.

We would be grateful if you could reconsider your rating in light of these responses and let us know whether any specific concern remains.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Closing Remarks

Dear Area Chair and Reviewers,

We thank the Area Chair and reviewers for their time and thoughtful feedback. Our paper introduces Fluxer, a new token mixer that generates image-dependent global convolution kernels in the spatial domain and applies them through FFT.

Our responses include kernel-generator comparisons with the surrounding backbone and training recipe held fixed, alongside comparisons with MLLA, VSSD, and Spatial-Mamba, which share the same four-stage backbone framework and closely aligned training protocols. We also provide comparisons with recent methods across classification, detection, and segmentation, including additional Cascade experiments not included in the original submission, and translation tests measuring prediction consistency and accuracy.

We hope these clarifications and supporting results have resolved the substantive concerns raised by all three reviewers. We would be grateful if they could be taken into account in the final evaluation of our paper.

Thank you again to the Area Chair and all reviewers for your time and effort.

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.