acceptodds
Under review as a conference paper at ICLR 2027

Dynamic-IsoCLIP: Cross-Modal Prior Spectral Routing for Instance-Aware CLIP Adaptation

Abstract

We propose Dynamic-IsoCLIP, a supervised and parameter-efficient framework for adapting frozen CLIP representations to image-to-image retrieval and base-to-new classification. Existing IsoCLIP-style methods improve intra-modal alignment through a discrete topological surgery, namely fixed hard truncation of the CLIP projector spectrum, but apply the same manually selected mask to every input. To address this limitation, Dynamic-IsoCLIP introduces Cross-Modal Prior Spectral Routing (CMPSR), which replaces the binary mask with a continuous, image-conditioned mask over a spectral basis constructed offline from the frozen image and text projectors. The text projector supplies an offline cross-modal geometric prior, while a lightweight MLP router selects spectral directions using only image features at inference, avoiding costly online cross-modal attention. Prototype supervision breaks the routing symmetry, and representation distillation and spectral sparsity act as semantic anchors and training stabilizers for the adaptation. Experiments on eleven benchmarks show that Dynamic-IsoCLIP achieves the highest image-to-image retrieval mAP among the evaluated supervised baselines. Base-to-new evaluation further shows improved performance over the original CLIP reference on five datasets under the specified text-classifier protocol, indicating that adaptive spectral routing can improve downstream discrimination while retaining unseen-class recognition within the evaluated setting.

Primary area:
foundation or frontier models, including LLMs
Keywords:
CLIP Adaptation, Spectral Routing, Cross-Modal Prior

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

This paper proposes Dynamic-IsoCLIP to replace the fixed hard spectral truncation in IsoCLIP with soft mask. The method constructs spectral basis from the frozen CLIP image and text projectors and trains a MLP to predict the contribution of each spectral direction, and only image features are required at inference time. The router is trained using class prototypes derived from frozen CLIP representations, together with a distillation loss and a mean-mask regularizer. The authors evaluate the approach on eleven image-to-image retrieval benchmarks, where test images serve as queries and training images form the gallery. The reported retrieval improvements are substantial. Additional experiments examine mask initialization, loss components, and base-to-new classification.

Critical strengths
  1. The motivation is reasonable. Soft mask is obviously better than fixed spectral basis. The proposed method is sample.
  2. The reported retrieval improvements are substantial.
Critical weaknesses
  1. Incremental contribution and insufficient isolation of dynamic routing. The contribution appears incremental, and the paper does not sufficiently establish that dynamic routing itself is the key source of the observed performance gains. There is no matched global learnable soft mask baseline using the same spectral basis and training objective.
  2. The baselines are too weak. The authors should train some method base on IsoCLIP, which is training-free. It is necessary to use supervision to train IsoCLIP as direct comparison. A fair comparison should include a supervised IsoCLIP baseline, where the fixed IsoCLIP mask is kept unchanged but the same prototype supervision, distillation loss, and sparsity regularization are used.
  3. Unclear contribution of the cross-modal prior. The additional value of the text projector in constructing the spectral basis is not sufficiently validated. Since the proposed spectral basis is derived from the frozen image and text projectors, the paper does not isolate how much the text information contributes beyond the image representation.
Questions for authors

None

3 comments

W 1. The central contribution is image-conditioned spectral routing over a frozen cross-modal basis. A global learnable soft mask assigns the same spectral weights to every image, whereas Dynamic-IsoCLIP predicts an input-dependent mask, allowing individual images to emphasize different spectral directions. This distinction goes beyond replacing hard truncation with continuous weights.

To directly isolate this contribution, we added a global learnable soft-mask baseline using the same spectral basis, supervision, loss functions, teacher features, class prototypes, optimizer, and training schedule (no-prior initialization). The following results report retrieval mAP (%) as mean ± standard deviation:

DatasetGlobal learnable soft maskDynamic-IsoCLIP
EuroSAT54.99±0.1295.88±0.11
Food10152.32±0.0981.73±0.06
CUB32.41±0.1052.94±0.06

Values report retrieval mAP. Dynamic-IsoCLIP values are the no-prior results reported in Table 2. Both methods optimize continuous spectral weights under the same objective, yet Dynamic-IsoCLIP achieves higher mAP by 40.89, 29.41, and 20.53 percentage points, respectively. Because a shared static vector rescales all images identically, it cannot adapt to sample-specific visual attributes and is easily over-constrained by the global sparsity penalty. These controls confirm the contribution of instance conditioning beyond global spectral-weight optimization.

W 2. A supervised fixed-mask baseline requires specifying which parameters remain trainable. In our architecture, the CLIP encoders, projection matrices, and cross-modal spectral basis are frozen, and the router is the only trainable representation component. Class prototypes are computed from frozen teacher features. Therefore, fixing the binary IsoCLIP mask leaves no trainable representation parameters: the task and distillation losses cannot change the representation, and the sparsity term is constant.

Under this architecture, label supervision yields two relevant controls:

  • Fixed IsoCLIP mask with supervised prototypes (IsoCLIP-NCM, the results on EuroSAT / Food101 / CUB are 52.37% / 45.74% / 39.03% mAP, respectively). The binary IsoCLIP mask H remains unchanged, and supervision is used only to construct class prototypes over the resulting fixed representations. This is the meaningful fixed-mask supervised reference.
  • Supervised global soft-mask control (Global-IsoCLIP). To enable optimization under the same prototype supervision, distillation loss, sparsity regularization, optimizer, and training schedule, we replace the fixed binary mask with one trainable soft spectral-weight vector shared by all images. The numerical comparison is provided in our response to Concern 1.

Our central innovation is image-conditioned spectral routing over a frozen cross-modal basis. The global control learns one spectral-weight vector for the entire dataset, whereas Dynamic-IsoCLIP predicts input-dependent weights for each image. Thus, the comparison isolates the benefit of instance conditioning beyond label supervision or global optimization of spectral weights.

W 3. The text projector contributes structurally through the construction of the cross-modal spectral basis, rather than through online text inputs to the router. This structural prior is distinct from mask initialization: the no-prior variant removes the IsoCLIP truncation initialization but retains the cross-modal directions 𝐔.

To isolate how much the text projector contributes beyond image-only representations or arbitrary orthogonal coordinates, we compared 𝐔 with an image-only PCA basis and a random orthogonal basis under the exact same router and no-prior training protocol. The following results report retrieval mAP (%) as mean ± standard deviation:

DatasetCross-modal basisImage-only PCARandom orthogonal
EuroSAT95.88±0.1191.48±0.8392.45±0.82
Food10181.73±0.0676.64±0.5379.17±0.70
CUB52.94±0.0646.52±0.6249.86±0.73

The cross-modal basis exceeds image-only PCA by 4.40–6.42 percentage points and the random orthogonal basis by 2.56–3.43 points. Geometrically, 𝐟∈ℝ⁷⁶⁸ has a higher dimension than the projected embedding (d=512). Image-only PCA in ℝ⁷⁶⁸ prioritizes high-variance domain/background directions that lie outside the subspace spanned by the image projector 𝐖ᵢ. In contrast, 𝐔 , the image-side left singular directions of 𝐖ᵢᵀ 𝐖ₜ ,spans a cross-modal subspace within row(𝐖ᵢ) and aligns the coordinates with multimodal semantic axes, yielding a superior structural basis at zero additional inference cost.

The rebuttal solved most of my concern, so I decide to raise my rating.

Dear Reviewer ZUjg,

Thank you very much for updating your score and for acknowledging our rebuttal efforts. We deeply appreciate the time and expertise you dedicated to evaluating our work.

Best regards.

Authors

Reviewer_q9naOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates parameter-efficient CLIP adaptation for image-to-image retrieval, arguing that the fixed spectral truncation used in IsoCLIP cannot account for differences across individual images. The authors propose Dynamic-IsoCLIP, which introduces Cross-Modal Prior Spectral Routing (CMPSR) to replace the fixed binary mask with an image-conditioned soft mask. The method constructs a spectral basis from frozen image and text projectors and uses a lightweight MLP to adaptively weight spectral components. Prototype-based supervision guides the routing process, while representation distillation and sparsity regularization constrain the adapted features.

The method is evaluated on eleven image-to-image retrieval benchmarks and five base-to-new classification datasets. The reported results show improvements over the compared baselines across all eleven retrieval benchmarks. Additional experiments examine spectral mask initialization, loss component ablations, and sample-wise routing behavior, providing empirical evidence for the proposed method.

Critical strengths
  1. The method is simple and computationally practical. Dynamic-IsoCLIP introduces a lightweight MLP router while keeping the CLIP backbone and spectral basis frozen. The image-conditioned masks allow different inputs to use different spectral directions without requiring online text features or cross-modal attention. This makes the approach relatively easy to integrate into image-to-image retrieval systems.
  2. The experimental results are promising across different datasets. Dynamic-IsoCLIP achieves the highest retrieval mAP among the compared methods on all eleven benchmarks, including fine-grained recognition and scene classification datasets. The ablation results in Table 3 also show that both distillation and sparsity regularization contribute to the performance. These results suggest that the proposed adaptation strategy is effective, even though the individual contributions of dynamic routing and the cross-modal spectral basis remain insufficiently isolated.
Critical weaknesses
  1. The benefit of instance-aware routing is not clearly demonstrated. In Table 2, Dynamic-IsoCLIP is compared with training-free IsoCLIP, but the two methods differ in both supervision and mask generation. A global learnable soft mask trained with the same objective would be a more meaningful baseline. Without this comparison, it is difficult to determine whether the improvements come from image-dependent routing or simply from learning better spectral weights.
  2. The role of the cross-modal spectral basis remains unclear. Eq. (1) constructs the basis using both image and text projectors, but the subsequent routing in Eqs. (4)–(6) operates entirely on image features. The paper does not compare this basis with an image-only spectral basis or a random orthogonal basis. Since the cross-modal prior is central to the proposed method, the authors need to show that it provides an advantage over these simpler alternatives.
  3. The experimental comparisons do not fully isolate the proposed method's contribution. Table 1 compares Dynamic-IsoCLIP with several CLIP adaptation methods, but these approaches differ in their training objectives and adaptation mechanisms. Some baselines were originally developed for classification rather than image-to-image retrieval. In addition, the paper reports results from a single CLIP backbone without repeated-run statistics. Given the large improvements on datasets such as EuroSAT and Food101, stronger supervision-matched comparisons and variance reporting are needed to assess the reliability of the results.
  4. The claim of preserving CLIP's generalization ability is not sufficiently supported. Table 4 evaluates base-to-new classification on only five datasets and compares Dynamic-IsoCLIP with the original CLIP model. Although the results improve, the evaluation remains within the same datasets and does not examine cross-domain transfer or image-text retrieval. Since the proposed spectral routing modifies the original CLIP embedding geometry, it is unclear whether the adapted representations retain the cross-modal alignment that makes CLIP useful beyond the supervised retrieval task.
Questions for authors
  1. Can the authors provide a comparison with a global learnable soft mask trained using the same supervision and loss functions? This would help clarify whether the improvement comes from instance-aware routing or simply from optimizing spectral weights.
  2. How important is the cross-modal spectral basis constructed in Eq. (1)? Would the method achieve similar results using an image-only spectral basis or a random orthogonal basis? This comparison is important for understanding the contribution of the proposed cross-modal prior.
  3. Could the authors clarify whether all methods in Table 1 were evaluated under the same training data, supervision, and retrieval protocol? Also, are the reported results averaged over multiple runs? These details are important for assessing the fairness and reliability of the large performance gains.
  4. Does the learned spectral routing preserve the original image-text alignment of CLIP? The base-to-new results in Table 4 are encouraging, but they do not directly measure cross-modal retrieval performance. A comparison of image-text retrieval before and after adaptation would help determine whether the method improves image discrimination at the expense of cross-modal alignment.
1 comment

W&Q1. We added a global learnable soft-mask baseline (M = σ(a/τₘ) ∈ (0,1)ʳ) trained with identical supervision, loss, frozen cross-modal spectral basis, teacher features, class prototypes, data split, optimizer and schedule (no-prior initialization). The distinction is that the global baseline learns one spectral-weight vector shared by all images, whereas Dynamic-IsoCLIP predicts weights conditioned on each input image. Retrieval mAP (%), mean ± std:

DatasetGlobal learnable soft maskDynamic-IsoCLIP
EuroSAT54.99±0.1295.88±0.11
Food10152.32±0.0981.73±0.06
CUB32.41±0.1052.94±0.06

Both supervised methods optimize continuous spectral weights under the same objective, yet Dynamic-IsoCLIP outperforms the global learnable soft mask by 40.89, 29.41, and 20.53 percentage points. Because a static global vector rescales all samples identically, it cannot adaptively suppress instance-specific background noise or highlight fine-grained attributes, and is heavily constrained by the global sparsity regularizer when trained without prior initialization. These results confirm that the gains stem from instance-aware routing rather than optimizing global spectral weights.

W&Q 2. Although the router takes only image features 𝐟 online for efficient independent retrieval encoding, the text projector 𝐖ₜ shapes the routing coordinate system offline via the SVD of 𝚿=𝐖ᵢᵀ 𝐖ₜ=𝐔𝚺𝐕ᵀ. Under the exact same router architecture and training protocol (no-prior), we compared 𝐔 against an Image-only PCA basis and a Random Orthogonal basis. Retrieval mAP (%), mean ± std:

DatasetCross-modal basisImage-only PCARandom orthogonal
EuroSAT95.88±0.1191.48±0.8392.45±0.82
Food10181.73±0.0676.64±0.5379.17±0.70
CUB52.94±0.0646.52±0.6249.86±0.73

The cross-modal basis consistently outperforms Image-only PCA by +4.40% to +6.42% and Random Orthogonal bases by +2.56% to +3.43%. Geometrically, 𝐟∈ℝ⁷⁶⁸ has a higher dimension than the projected embedding (d=512). Image-only PCA in ℝ⁷⁶⁸ prioritizes high-variance domain/background directions that lie outside the subspace spanned by the image projector 𝐖ᵢ. In contrast, 𝐔 , the image-side left singular directions of 𝐖ᵢᵀ 𝐖ₜ ,spans a cross-modal subspace within row(𝐖ᵢ) and aligns the coordinates with multimodal semantic axes, yielding a superior structural basis at zero additional inference cost.

W&Q 3. All methods in Table 1 were trained and evaluated using the exact same training splits, label supervision, frozen CLIP ViT-B/16 features, gallery/query partitions, and mAP metric.

Our global soft-mask control plus basis ablation form strict baselines trained under the exact same 𝓛 objective. Specifically, the Random Orthogonal Basis + Router variant serves as an MLP-gated adaptation control trained under the same prototype supervision, distillation loss, and sparsity regularization, with a random orthogonal basis replacing the cross-modal spectral basis. Dynamic-IsoCLIP outperforms this control by 2.56–3.43 mAP points.

All results in the paper are averaged over five random seeds. The statement in the conclusion"does not report repeated-run uncertainty"only refers to missing standard deviations in the initial submission. We removed this sentence and added std values to tables in revision. We additionally report the retrieval mAP (%) as mean ± standard deviation for the two variants of Dynamic-IsoCLIP:

InitializationAircraftEuroSATFood101
No prior27.91 ± 0.1095.88 ± 0.1181.73 ± 0.06
IsoCLIP prior28.23 ± 0.1296.31 ± 0.1381.95 ± 0.09

The no-prior variant also performs close to the prior-initialized variant, supporting effective instance-conditioned routing without requiring IsoCLIP-mask initialization.

W&Q 4. We clarify how our framework preserves image-text alignment both theoretically and empirically:

Structural & Loss Anchoring: Unlike unrestricted fine-tuning of 𝐖ᵢ, CMPSR only reweights coordinates within the existing cross-modal subspace 𝐔 of 𝐖ᵢᵀ 𝐖ₜ, and explicitly anchors the routed image embedding 𝐳̃ₙ to the frozen CLIP teacher 𝐳ₙᵗᵉᵃ via 𝓛_distill=1−cos(𝐳̃ₙ,𝐳ₙᵗᵉᵃ)=½ ‖𝐳̃ₙ−𝐳ₙᵗᵉᵃ‖₂².

Direct Cross-Modal Verification in Table 4: In our Base-to-New experiments, classification is performed by computing cosine similarity between the routed image embeddings and frozen CLIP text embeddings ("a photo of a [CLASS]"). Even on Unseen (New) classes never seen by the router during training, Dynamic-IsoCLIP improves cross-modal classification accuracy over original CLIP across all 5 datasets.

Furthermore, evaluating Text-to-Image (T2I) retrieval using class prompts on EuroSAT / Food101 / CUB shows that Dynamic-IsoCLIP improves T2I mAP from 52.89% / 80.54% / 47.05% to 54.37% / 87.92% / 47.88% over original CLIP, confirming that cross-modal collapse does not occur.

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

The paper introduces a new method for image retrieval that builds on the existing algorithm IsoCLIP. IsoCLIP aims to fix pretrained Vision Language models like CLIP by applying a fixed binary mask to hard-truncate CLIP’s projection spectrum. The author then states that it's a global mask that may suppress useful visual evidence needed for more fine-grained tasks. To solve this, the author introduced Dynamic-IsoCLIP, which uses CMPSR (Cross-Modal Prior Spectral Routing), keeping CLIP’s spectral basis frozen while employing a lightweight 2-layer MLP router to predict a continuous, sample-specific soft mask.The router learns these masks from labelled images, while the rest of CLIP remains frozen.

Critical strengths
  1. The paper is well written and highlights a critical issue in ISOCLIP.
  2. The results demonstrate the utility of the proposed method.
  3. Ablations are thorough and demonstrate the utility of the proposed components.
Critical weaknesses

L1: The main contribution of this paper is to introduce a dynamic and sample-dependent mask over the spectral basis. However, as acknowledged by the authors, there are limitations regarding the lack of a learned global soft-mask baseline. This omission directly challenges the paper's main claim, as it remains unclear whether the accuracy gain is primarily due to learning the mask or simply to replacing the hard mask with a soft one. As instance-aware masking is central, the author shall compare these masking techniques under the same training objective. They could compare a learned global soft mask, a class-specific mask, a fixed global mask, and the proposed image-specific mask to quantify the algorithm's accuracy gain.

L2: As an extension to training-free IsoCLIP, this paper requires labelled per-dataset training, whereas IsoCLIP does not, thereby removing the main advantage of IsoCLIP. Training time and re-encoding cost are not reported, and scalability is provided only by an analytical estimate. Therefore, for a true comparison, training time for each baseline across all datasets shall be reported.

L3: The paper does not specify the number of random seeds used for the proposed MLP router, as it is trained stochastically; there will be some uncertainty along the seeds. Please report the mean ± standard deviation of the proposed model over multiple independent runs. This is important for conclusions drawn from the reported benefits of IsoCLIP-prior initialisation on Food101 (+0.22), Aircraft (+0.32), and EuroSAT (+0.43), where the differences are relatively small.

L4: IsoCLIP motivates using the SVD basis for hard truncation, where singular-value ordering determines which directions are removed. In the supervised part, Σ is never used, and the no-prior variant ignores the ordering. So does the benefit of the cross-modal basis carry over to a learned soft router? Please report a control with a random orthogonal basis and an image-only PCA basis.

Questions for authors

Please refer to the limitation points (L1-L4) in the weakness section above, especially the last line that involves a direct question the authors can aim to rebut.

3 comments

L1. The key distinction is between learning continuous spectral weights and conditioning those weights on individual images. Our contribution is the latter: Dynamic-IsoCLIP learns an input-conditioned mask, 𝓜, whereas a learned global soft mask uses a single trainable vector 𝐯 shared by all images.

We trained a global soft-mask baseline using the exact same frozen cross-modal spectral basis, training data and split procedure, teacher features, class prototypes, loss functions, optimizer, and 20-epoch schedule. Both configurations use ViT-B/16 and no-prior initialization. Retrieval mAP (%), mean ± std:

DatasetLearned global soft maskDynamic-IsoCLIP
EuroSAT54.99±0.1295.88±0.11
Food10152.32±0.0981.73±0.06
CUB32.41±0.1052.94±0.06

Dynamic-IsoCLIP exceeds the learned global soft-mask baseline by 40.89, 29.41, and 20.53 percentage points, respectively. For the global soft-mask baseline, a single static vector 𝐯 rescales all images identically. It cannot adaptively alter sample-specific neighborhood rankings in cosine space, and optimizing it from a no-prior initialization under the uniform sparsity penalty 𝓛 suppresses spectral directions globally without instance selectivity.

Regarding a class-specific mask: in image-to-image retrieval, query class labels are unknown at inference time (test labels are used solely to evaluate ranking relevance), meaning ground-truth class-specific masks cannot be applied to test queries without label leakage, while pseudo-labeled class masks cannot capture intra-class variations in pose, scale, and background. Conditioning directly on the visual feature 𝐟 allows Dynamic-IsoCLIP to remain label-free at inference while adapting to sample-level visual evidence.

L2. Our contribution is supervised, image-conditioned spectral routing over a frozen cross-modal basis. We measure training latency on RTX 5060 Ti (16 GB), times in seconds:

MethodEuroSATFood101CUB
Linear Probe9.5337.053.58
CLIP-AST2.569.520.91
MergeTune-style6.3323.052.42
Dynamic-IsoCLIP3.5613.161.28

Measurements use cached frozen-encoder features, excluding one-time initial feature extraction and model loading. Our timing includes validation and checkpoint saving; even so, Dynamic-IsoCLIP finishes 20 epochs in 1.28–13.16 s, faster than Linear Probe and MergeTune-style, and much faster than prompt-tuning that backpropagates through encoders.

Only the lightweight router (0.16M params) is trained, avoiding repeated image encoding. Full gallery refresh takes 0.022 / 0.089 / 0.009 s on EuroSAT / Food101 / CUB. We will provide the complete training-time comparisons across all baselines and datasets in Appendix B.

L3. We test two initialization schemes with five independent runs. Retrieval mAP (%), mean ± std:

InitializationAircraftEuroSATFood101
No prior27.91 ± 0.1095.88 ± 0.1181.73 ± 0.06
IsoCLIP prior28.23 ± 0.1296.31 ± 0.1381.95 ± 0.09

IsoCLIP-prior initialization achieves higher mean mAP on all three datasets with very small standard deviations. The strong performance of the no-prior variant shows that effective instance-conditioned spectral routing can be learned purely from task supervision, without requiring IsoCLIP's manual truncation rule.

L4. In the no-prior variant, neither singular-value magnitudes (𝚺) nor their ordering initialize the mask. However, the basis vectors 𝐔 still come from the cross-modal operator 𝐓, defining the coordinate axes along which the router reweights image features. Under identical router architecture and training protocol, we compare 𝐔 with image-only PCA and random orthogonal bases. Retrieval mAP (%), mean ± std:

DatasetCross-modal basisImage-only PCARandom orthogonal
EuroSAT95.88±0.1191.48±0.8392.45±0.82
Food10181.73±0.0676.64±0.5379.17±0.70
CUB52.94±0.0646.52±0.6249.86±0.73

The cross-modal basis exceeds image-only PCA by 4.40--6.42 percentage points and the random orthogonal basis by 2.56--3.43 points. Geometrically, the pre-projection features have a higher dimension than the projected space (d=512). Image-only PCA ranks directions by image-feature variance and tends to retain high-variance background/domain directions that lie outside the projector's active row space; this also explains why PCA underperforms the isotropic random basis on all three datasets. In contrast, U consists of the image-side left singular directions of Ψ = WᵢᵀWₜ, which incorporate the relationship between the frozen image and text projectors and align the routing coordinates with multimodal semantic axes. Since the basis is constructed offline, it introduces no additional online inference cost relative to the alternative bases.

Reviewer_kDhfOpenReview ↗

Reviewer response

Dear Authors,

I thank for the detailed rebuttal that clarifies my doubts. I maintain my original rating of clear acceptance. Good paper.

Dear Reviewer kDhf,

Thank you very much for taking the time to read our rebuttal and for your positive feedback and strong support of our work. We truly appreciate your constructive comments, which have helped us improve the paper.

Best regards,

Authors

Comments on the paper

1 comment

Thank you for identifying these presentation and proofreading issues. We have carefully addressed the presentation details and typographic errors in the updated manuscript.

Tables 1 and 2: We have reorganized the layouts and enlarged the text to reduce density. Table 1 now uses a clearer two-panel design, and redundant configuration columns in Table 2 have been streamlined into the caption for better readability.

Figure 3: The raster image has been replaced with a high-resolution vector PDF. We have significantly enlarged all axis labels, tick labels, panel titles, and the legend, while introducing a shared horizontal-axis title to ensure clear visibility.

Proofreading: We have conducted a thorough final pass of the manuscript, correcting the missing space ("regularizer. All" in Section 4.1), the capitalization ("first" instead of "First"), and ensuring overall formatting consistency.

With these presentation details fully resolved, and noting that no outstanding technical or methodological concerns remain from the previous discussion, we believe the revised manuscript now clearly communicates our scientific contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.