Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Semi-supervised learning (SSL) has achieved promising results in medical image segmentation by leveraging unlabeled data via consistency regularization. However, in multi-modal settings, heterogeneous modality characteristics and imbalanced reliability make effective consistency learning challenging. Existing methods typically impose prediction- or representation-level consistency on entangled features, which may suppress modality-specific cues and limit the exploitation of complementary cross-modal information. To address this limitation, we propose HDCL, a hierarchical disentangled consistency learning framework to promote reliable and robust semi-supervised multi-modal medical image segmentation. Specifically, Disentangled Representation Learning (DRL) explicitly decomposes multi-modal features into shared and modality-specific representations, enabling targeted modeling of invariant semantics and modality-specific cues. We then propose Shared-Specific Dual-path Decoding (SSDD), which simultaneously decodes shared, specific, and fused representations to produce diverse modality-conditioned predictions. Furthermore, Multi-level Prediction Consistency Learning (MPCL) is proposed to enforce prediction consistency across different semantic spaces, allowing more effective utilization of unlabeled data without compromising modality-specific characteristics. Extensive experiments on two benchmarks demonstrate that our method consistently outperforms existing methods across different annotation settings. The source codes will be available at https://github.com/FeixiangZhou/HDCL.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0554_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/FeixiangZhou/HDCL

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhoFei_Hierarchical_MICCAI2026,
        author = { Zhou, Feixiang AND Gao, Zhuangzhi AND Xie, Jianyang AND Li, Kun AND Wang, Fu AND Chen, Long AND Jiang, Zheheng AND Zhao, Yitian AND Meng, Yanda AND Zhao, He AND Lip, Gregory Y. H. AND Zheng, Yalin},
        title = { { Hierarchical Disentangled Consistency Learning for Semi-supervised Multi-modal Medical Image Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The work is to focus on a predictive model with its ability to deal with multi-modalities given by labeled and unlabeled medical images. Three strategies, disentangled representation learning, shared-specific dual-path decoding, and multi-level prediction consistency learning were studied.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The proposed method has a clear advantage to the state-of-the-art approaches on the ISLES-2022 and BraTS-2019 datasets. Three strategies were studied, with ablation results to show their individual contributions.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The proposed method is considered an incremental effort. There are several components that contribute to good performance. However, one cannot be sure whether we should use all the three strategies as an integrated system or each of the strategies can be helpful in various scenarios.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    • The title said the approach is for medical image segmentation. However, the methodology is mainly about the discussion on how to disentangle the representation to the shared and the modality-specific components. It means the novelty is on the general methodology rather than medical-related breakthroughs.

    • How exactly the proposed method can encourage the shared representation between different modalities to own as much information as possible? Through Eg. 1, 2, or 3? Or this issue is not important?

    • The balancing coefficients in the final loss function contain $alpha$ and $beta$. Does it mean that the rest of the components need no weight balance? (Page 6, right before Section 3.)

    • In the qualitative analysis, do we have any medical insight to describe the segmentation results? Or just better performance near the boundaries?

    • In the ablation study, the author(s) discussed the contribution from different strategies: DRL, SSDD, MPCL. Can we conclude that DRL is the most important strategy among all? Which one should receive the highest credit among all? What should we emphasize in this study, or it is an integrated efforts from all strategies? According to the results, DRL may not be the most influential one? In fact, each strategy should have its own contribution and should be shown along some results to further clarify each of their contributions.

    • According to the past studies, most reported numbers on those evaluation metrics, such as the ones on Table 1 may be with high variance. One should also include the variance (standard deviation) for reference in this study. On the other hand, the reported numbers in this study are based on some averaged results? How many trials to obtain the averages?

    • It is recommended to include a small paragraph or a table to summarize the notations used in this work. Quite a few of notations (e. q. , W_Q, W_K, W_V) that were used without their definitions to be mentioned.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Technical depth and clear presentation. The technical contribution is non-trivial. The presentation provides enough insight for readers to understand the contribution of the work.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper proposes a semi-supervised multi-modal medical image segmentation framework, that combines disentangled representation learning and hierarchical consistency learning to better exploit unlabeled multi-modal data. The method consists of three main components. First, the Disentangled Representation Learning (DRL) module separates features into shared and modality-specific representations using a shared encoder and modality-specific encoders, while introducing losses to align shared representations and reduce redundancy between shared and specific features. Second, the Shared-Specific Dual-path Decoding (SSDD) module produces predictions from shared, modality-specific, and fused semantic spaces; in particular, modality-specific features are enhanced through cross-attention before decoding, and shared/specific features are further fused for final prediction. Third, the Multi-level Prediction Consistency Learning (MPCL) module imposes both intra-space and inter-space consistency across different decoding branches, using fused predictions as pseudo-supervision for unlabeled data. Overall, the paper’s key idea is to preserve modality-specific information while still enforcing cross-modal consistency, which is a meaningful direction for semi-supervised multi-modal segmentation.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    This is an interesting framework for semi-supervised multi-modal medical image segmentation. A key strength is that it explicitly disentangles shared and modality-specific representations and then performs hierarchical consistency learning across shared, specific, and fused semantic spaces. This is a meaningful design because multi-modal segmentation naturally involves both modality-invariant semantics and modality-dependent cues, and the proposed framework is well motivated to preserve both instead of enforcing consistency on entangled features only. The experimental results are also strong: on both ISLES-2022 and BraTS-2019, the proposed method consistently outperforms prior baselines under different low-label settings. The gains are not limited to Dice but also appear in Jaccard, 95HD, and ASD, suggesting improvements in both overlap quality and boundary accuracy. The ablation study is also well aligned with the method design, showing that DRL, SSDD, and MPCL each contribute positively, while the qualitative examples and t-SNE visualization further support the effectiveness of the learned representations.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The main weakness is that the method is only clearly defined for two modalities. This is especially evident in Eq. (3), where the cross-attention mechanism is formulated only between two modality-specific features. It is unclear how the method would generalize to three or more modalities, including how the cross-attention, fused representation, and consistency losses would be defined in that case. Since many practical medical imaging settings involve more than two modalities, this limits the clarity and potential generality of the proposed framework. Minorly, the novelty should be positioned more carefully: disentangling shared and modality-specific features is not entirely new, so the more distinctive contribution appears to be the combination of disentanglement with hierarchical multi-space consistency learning for semi-supervised segmentation and medical images.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    i). The paper would be stronger if the authors clarify how the framework generalizes beyond the two-modality case, especially the cross-attention design in Eq. (3) and the fused consistency scheme.

    ii). Since the method is motivated by practical multi-modal settings, a brief discussion of computational cost and scalability would also be useful, especially given the multiple encoders, decoders, and consistency branches.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    My main concern is the lack of discussion on extension to more than two modalities. However, I do not view this as a fatal weakness for the current submission, because the paper already demonstrates a useful and effective method for the two-modality setting. The core idea is good, the design is reasonable, and the empirical results are strong enough to support acceptance.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper presents HDCL (Hierarchical Disentangled Consistency Learning), a framework for semi-supervised multi-modal medical image segmentation that addresses a fundamental tension in existing approaches: how to enforce meaningful consistency regularization across modalities without collapsing modality-specific information that may be clinically relevant. The core intellectual contribution is the recognition that prior consistency-based SSL methods operate on entangled feature representations, meaning shared semantics and modality-specific characteristics are mixed together in a single latent space. When consistency losses are applied to these entangled features, the optimization pressure tends to homogenize representations across modalities, inadvertently suppressing the very complementary information that makes multi-modal imaging valuable in the first place. HDCL addresses this through three tightly coupled components. First, Disentangled Representation Learning (DRL) explicitly separates multi-modal features into shared and modality-specific components using a shared encoder alongside modality-specific encoders. A cross-modal alignment loss (L1 distance between shared representations) encourages modality-invariant semantics, while a cosine similarity minimization loss discourages redundancy between shared and specific representations. This structured decomposition provides the foundation for all subsequent components. Second, Shared-Specific Dual-path Decoding (SSDD) builds three parallel decoding branches operating in distinct semantic spaces: a shared branch decoding modality-invariant representations, a specific branch using cross-attention to allow each modality to selectively incorporate complementary cues from the other while retaining its own characteristics, and a fused branch that concatenates shared and attention-enhanced specific features. Feature perturbation is applied in the fused space to generate a second prediction per decoder, yielding six total predictions per training sample. This multi-space decoding is designed to produce prediction diversity that can be systematically exploited by the consistency framework. Third, Multi-level Prediction Consistency Learning (MPCL) leverages these diverse predictions through both intra-space and inter-space consistency. Intra-space consistency is applied to the shared and fused branches (but deliberately withheld from the specific branch to preserve modality-specific cues), while inter-space consistency uses the fused decoder’s predictions as pseudo-labels to supervise the shared and specific branches. This asymmetric design reflects an important design choice: the fused space, integrating both shared and specific information, is treated as the most reliable prediction source and used to guide the other branches. The framework is evaluated on ISLES-2022 (ischemic stroke, DWI+ADC) and BraTS-2019 (glioma, T1ce+T2) under 5% and 10% labeled data settings, demonstrating consistent state-of-the-art performance against six competing methods.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Well-motivated problem formulation with a clear gap identification. The paper identifies a genuine and underappreciated limitation of prior work: that consistency regularization applied to entangled multi-modal features creates a conflict between cross-modal alignment and preservation of modality-specific cues. This is articulated clearly in the introduction and supported by both the t-SNE visualizations (Fig. 3b) and the ablation results. The motivation is not merely asserted but structurally embedded in the design of every component, giving the paper a coherent internal logic. 2.Principled disentanglement strategy with complementary losses. The DRL component combines alignment and disentanglement objectives in a principled way. Using L1 distance for shared representation alignment is well-suited to encouraging distributional overlap across modalities, while cosine similarity minimization for shared-specific orthogonality is an appropriate inductive bias for promoting independence in the representation space. The combination of these two objectives, pulling shared representations together across modalities while pushing shared and specific representations apart within each modality, is technically sound and motivated by established multi-modal disentanglement literature (e.g., Wang et al. [14], Wang et al. [17]). 3.The asymmetric treatment of semantic spaces in MPCL is a thoughtful design choice. Excluding the modality-specific branch from intra-space consistency regularization, while using the fused branch as pseudo-label teacher for inter-space consistency, reflects careful reasoning about reliability hierarchies among the different prediction spaces. This avoids the naive approach of simply applying uniform consistency pressure everywhere, and the ablation in Table 3 provides empirical support that combining intra- and inter-space losses is complementary rather than redundant. 4.Thorough and well-structured ablation study. Table 2 evaluates each component incrementally across both datasets, demonstrating monotonic and meaningful improvement as DRL, SSDD, and MPCL are added. Table 3 further isolates the contributions of intra- and inter-space consistency. The results are consistent across datasets with different modality characteristics (stroke vs. glioma), lending confidence that the gains are not dataset-specific artifacts. The ablation design is logically ordered and sufficient to attribute performance gains to specific design decisions. 5.Strong quantitative performance across all evaluation settings. The margin over the strongest baseline (SMMS) is substantial on ISLES-2022 at 5% labeled data (3.38% Dice improvement), and improvements are consistent across four evaluation metrics (Dice, Jaccard, 95HD, ASD) on both datasets and both annotation settings. Improvements in boundary-sensitive metrics (95HD, ASD) are particularly notable, suggesting the disentangled representation learning is contributing to better structural delineation rather than simply inflating volumetric overlap scores. 6.Feature perturbation in the fused space is a pragmatic augmentation strategy. Introducing stochastic perturbation (noise or dropout) to the fused representation to generate a second prediction is a lightweight but effective way to encourage consistency under feature-level variation, analogous to data augmentation in the prediction space. This is straightforwardly motivated by perturbation-based SSL literature and adds prediction diversity without requiring architectural redundancy.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.Insufficient justification for the choice of consistency direction in MPCL. The paper designates the fused decoder’s predictions as pseudo-label teachers for both the shared and specific branches in inter-space consistency, arguing that the fused space is “more reliable.” However, this reliability assumption is not empirically validated beyond the final performance comparison. There is no analysis showing that fused-space predictions have higher confidence, lower entropy, or better calibration than shared-space predictions at various stages of training. This is particularly relevant in early training when the fused features may themselves be unreliable. An analysis of prediction confidence or calibration across the three branches over training would substantially strengthen this design choice. 2.The cross-attention mechanism in the specific branch lacks ablation. The SSDD component introduces a cross-attention module to allow each modality’s specific representation to attend to the other’s specific representation before decoding. This is a non-trivial design choice with its own parameters and computational cost, and it is the key mechanism distinguishing the specific branch from a naive dual-decoder setup. However, there is no ablation isolating the contribution of this cross-attention module specifically. It is impossible to determine from the provided experiments whether the gains from SSDD stem from the multi-space decoding structure, the cross-attention interaction, or both. Replacing cross-attention with simple concatenation or addition as a baseline would clarify this. 3.Limited analysis of the disentanglement quality. The t-SNE visualization in Fig. 3(b) provides qualitative evidence that shared representations align across modalities and specific representations separate into modality-specific clusters. However, t-SNE is a dimensionality reduction technique sensitive to perplexity and initialization, and visual cluster separation in 2D projections does not guarantee meaningful disentanglement in the original feature space. Quantitative metrics such as mutual information between shared and specific representations, or a modality classification accuracy on the specific features (higher accuracy would indicate better modality-specificity), would provide more rigorous evidence that DRL achieves its stated objective. 4.The method is evaluated only on brain MRI with two modalities. Both benchmark datasets are brain lesion segmentation tasks with two modalities each. While these are standard benchmarks in the field, they share similar imaging characteristics (MRI) and anatomical domain (brain). The generalizability of HDCL to settings with more than two modalities (e.g., the four modalities available in BraTS but not used here), different anatomical regions, or modality combinations with larger distributional gaps (e.g., CT and MRI) is not demonstrated. The paper’s claim of general applicability to multi-modal medical image segmentation would be better supported by at least one experiment outside brain MRI. 5.The specific branch excludes the modality-specific space from intra-space consistency without fully exploring alternatives. The decision to exclude the specific branch from intra-space consistency is justified qualitatively (to preserve modality-specific cues), but no experiment tests what happens when intra-space consistency is also applied to the specific branch, or with reduced weight. It is possible that a soft constraint rather than complete exclusion could be beneficial. Without this baseline, the binary design choice is not fully justified empirically. 6.Computational cost and inference complexity are not discussed. HDCL introduces three encoders (one shared, two modality-specific), three decoders, a cross-attention module, and multiple consistency losses. Compared to baselines using simpler dual-encoder-decoder architectures, the parameter count and training cost are likely substantially higher. The paper reports using two H100 GPUs with 60k iterations but does not compare inference time, model size, or FLOPs against competing methods. For clinical deployment, where computational constraints may be significant, this information is essential to assess practical feasibility. 7.Pseudo-label threshold sensitivity is not analyzed. The confidence threshold τ = 0.95 used in MPCL is set without ablation. This is a critical hyperparameter: too high a threshold may exclude too many unlabeled samples (especially in the 5% setting where the model may be poorly calibrated early in training), while too low a threshold introduces noisy pseudo-labels. The impact of this hyperparameter choice on performance, especially under the most data-limited setting, deserves dedicated analysis.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper presents HDCL, a hierarchical disentangled consistency learning framework for semi-supervised multi-modal medical image segmentation. The recommendation to accept is driven primarily by the quality and coherence of the core intellectual contribution, the rigor of the experimental validation, and the strength of quantitative results, despite several real weaknesses that would benefit from author response or revision. Factors supporting acceptance: The paper identifies a genuine and underappreciated problem: existing consistency-based SSL methods operate on entangled multi-modal features, creating an inherent conflict between cross-modal alignment and preservation of modality-specific information. This framing is not merely rhetorical, it is structurally embedded in every component of the design. The three proposed modules (DRL, SSDD, MPCL) form a coherent and well-motivated pipeline where each component has a clear purpose and the interactions between them are principled rather than ad hoc. The quantitative results are strong and consistent. On ISLES-2022 under the 5% labeled setting, HDCL outperforms the strongest baseline (SMMS) by 3.38% Dice, with consistent improvements across all four evaluation metrics and both datasets. Gains in boundary-sensitive metrics (95HD, ASD) suggest the framework is genuinely improving structural delineation rather than inflating volumetric overlap scores, which is a more demanding and clinically relevant criterion. The ablation study is well-designed: Table 2 demonstrates monotonic improvement as each component is added, and Table 3 establishes that the intra- and inter-space consistency losses are complementary. These results hold across two datasets with different modality pairs and pathology characteristics, reducing the risk that gains are dataset-specific artifacts. The asymmetric design of MPCL, excluding the specific branch from intra-space consistency while using the fused branch as pseudo-label teacher, reflects thoughtful reasoning about reliability hierarchies among semantic spaces. This avoids the naive approach of applying uniform consistency pressure everywhere, and the empirical ablation supports the design choice. Factors that reduce confidence and should be addressed in rebuttal or revision: Several weaknesses temper the enthusiasm for a stronger score. The reliability assumption underlying MPCL’s inter-space consistency, that the fused decoder is a more trustworthy teacher, is not validated beyond final performance. An analysis of prediction calibration or confidence across branches over training would strengthen this design choice considerably. The cross-attention module within SSDD, a non-trivial and parameter-heavy component, is not ablated in isolation, making it impossible to attribute gains between the multi-space structure and the cross-attention interaction specifically. The disentanglement evaluation relies solely on t-SNE visualization, which is qualitative and sensitive to hyperparameters; quantitative disentanglement metrics would be more convincing. Evaluation is also limited to brain MRI with two modalities, and the generalizability of HDCL to larger modality sets, different anatomical domains, or cross-modality settings (e.g., CT+MRI) remains an open question. Finally, the paper provides no comparison of computational cost, parameter count, or inference time against baselines, which is a meaningful omission given the added complexity of three encoders, three decoders, and cross-attention. Overall: The core idea is sound, the motivation is clear, the ablations are informative, and the empirical gains are substantial and consistent. The weaknesses are real but none individually undermines the central contribution. This warrants acceptance independent of rebuttal, though the authors are encouraged to address the calibration analysis, cross-attention ablation, and computational cost discussion in a revision.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We sincerely thank the Area Chairs and reviewers for their valuable time and constructive feedback. We will carefully address the comments and further improve the paper in the revised version.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The reviewers provide generally positive evaluations of this paper (scores: 3, 5, 5), with strong support for the overall methodological design and empirical performance, alongside some concerns regarding novelty, clarity, and generalization. Two reviewers highlight the well-motivated framework, coherent design, and consistent improvements across datasets, while another raises questions about incremental novelty, reproducibility, and the need for deeper analysis of individual components. Overall, I believe the paper makes a solid and meaningful contribution, and the strengths outweigh the identified weaknesses. I therefore recommend acceptance. The authors are encouraged to address the raised concerns in the camera-ready version, particularly by clarifying the novelty positioning, strengthening experimental evidence, and further discussing generalization and scalability.



back to top