acceptodds
Under review as a conference paper at ICLR 2027

Cross-Domain Few-Shot Segmentation via Dual-level Distillation on SAM3

Abstract

Cross-Domain Few-Shot Segmentation (CD-FSS) aims to segment target objects in unseen domains using only a few annotated references. Recent Segment Anything Model (SAM)-based methods require less task-specific training but rely on pairwise visual matching, whose reliability decreases when reference and target appearances differ, leading to substantial variation in per-target IoU. SAM3 supports promptable concept segmentation and provides a semantic prompt space for learning a reusable category prototype. This prototype produces stable predictions across targets but cannot fully represent the different appearances. Based on these insights, we propose Dual-level Distillation on SAM3 (D-SAM3), a lightweight framework that uses self-distillation at complementary category and image levels. (i) Category Prototype Distillation transfers GT-box-prompted predictions into a reusable category prototype; (ii) Visual Residual Distillation calibrates reference-target visual correspondence separately for each category and expresses the resulting visual prediction as a residual relative to the prototype-based prediction; and (iii) Reliability-aware Fusion controls this residual based on correspondence reliability and uncertainty of the prototype-based prediction. Together, they produce a category-stable yet target-adaptive representation. With SAM3 frozen, D-SAM3 optimizes only approximately 8.2K parameters per category. Extensive experiments demonstrate strong CD-FSS performance, robustness to reference selection, and generalization across tasks and semantic granularities.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cross-Domain Few-Shot Segmentation

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_sqH1OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies cross-domain few-shot segmentation (CD-FSS) and identifies a stability-adaptability trade-off in existing approaches. Pairwise reference-target visual matching can adapt to the appearance of an individual target but is sensitive to the selected reference and becomes unreliable under large appearance differences. A category prototype is more stable across target images but may fail to represent their diverse appearances.

To address this trade-off, the paper proposes D2-SAM3, a lightweight framework built on a frozen SAM3. Category Prototype Distillation (CPD) uses predictions produced from ground-truth box prompts on the annotated references as teachers for learning a reusable category prototype. Visual Residual Distillation (VRD) learns category-specific calibration of visual correspondence and expresses the visual prediction as a residual relative to the prototype prediction. Reliability-aware Fusion (RAF) modulates this residual using correspondence reliability and uncertainty in the prototype-based prediction. Only approximately 8.2K parameters are optimized per category.

Critical strengths
  1. The distinction between stable but incomplete category prototypes and adaptive but unreliable pairwise correspondence is supported by per-target IoU distributions and reference-selection experiments. Modeling the correspondence signal as a residual rather than as the primary prediction is both technically sensible and well matched to this diagnosis.

  2. D2-SAM3 adapts only about 8.2K parameters per category while keeping SAM3 frozen, and it performs strongly across four substantially different target domains. The component ablations, reference sensitivity analysis, efficiency measurements, task-transfer experiments, and backbone-controlled comparisons collectively provide credible evidence that the gains are not merely due to one favorable benchmark or a larger trainable model.

Critical weaknesses
  1. The reusable prototype is defined in SAM3’s concept-prompt space, and the method learns a new prototype and calibration parameters for every category. The paper does not establish whether the main principle transfers to other promptable models or how adaptation/storage scale to hundreds or thousands of categories. This limits the generality of the claimed framework despite the small per-category parameter count.

  2. CPD relies on a teacher produced by a ground-truth reference box, while VRD and RAF use correspondence reliability and prototype uncertainty to determine the correction. The paper does not test noisy masks/boxes or demonstrate that the reliability and uncertainty scores are calibrated predictors of actual error. These omissions are important because errors in the few available references could be amplified into a shared category prototype.

Questions for authors
  1. How does D2-SAM3 behave when reference masks or their derived boxes are noisy? Please report controlled mask erosion/dilation, boundary noise, box jitter, and missing-part experiments. Because CPD distills a teacher derived from these annotations, this result would materially affect my confidence in the method’s robustness.

  2. Can the authors quantify whether the correspondence reliability and prototype uncertainty used by RAF are calibrated? For example, please report their correlation with per-target visual/prototype prediction errors, reliability diagrams, or performance when either score is replaced by a learned or random ranking. This would validate the interpretation underlying RAF rather than only its aggregate mIoU benefit.

  3. In the 5-shot setting, why are only the two references with the highest leave-one-out utility retained? Please compare against using all references, random selection of two references, and a simple learned or similarity-weighted aggregation. Please also clarify the computational scaling for larger values of K.

  4. What is the total adaptation time and storage requirement when the number of categories becomes large? Can prototypes be amortized, initialized from text, or shared across related categories? This affects the practical meaning of the “lightweight” claim.

4 comments
AuthorsOpenReview ↗

Response to Reviewer sqH1

We thank the reviewer for recognizing our analysis of the stability–adaptability trade-off, the rationale for visual residual correction, and the extensive empirical evidence. We appreciate the constructive suggestions and address the concerns below.

Weakness 1a [Choice of SAM3 and transferability] Our motivation requires a segmentation model that combines a reusable semantic prompt space with visual correspondence. SAM3 is particularly well suited to this formulation: its promptable concept segmentation supports text and box-specified image exemplars within one model. Its large-scale concept supervision includes approximately 4 million distinct phrases and 52 million masks in SA-Co/HQ, alongside extensive synthetic training data. These capabilities allow us to distill box-prompted predictions into reusable category prompts while keeping the backbone frozen. SAM1/2 do not provide the same native text-conditioned concept-segmentation interface, so our implementation cannot be transferred directly.

Weakness 1b & Question 4 [Adaptation cost and scalability] The approximately 2-minute preparation time in Tab.6 covers three ISIC categories, with approximately 35 seconds per category (refer to revised Sec.5.6). Even we have hundreds of categories in practical scenarios, each new category prototype can be initialized from its textual embedding, but the learned prototype is category-specific and is not directly reused across categories. Each category can be incorporated through a very short 35-second adaptation, supporting rapid extension. In the scope of CD-FSS, four widely adopted benchmarks contain at most six categories.

Weakness 2 & Question 1 [Robustness to reference annotation noise] Standard CD-FSS protocols provide ground-truth reference masks and do not contain annotation errors. Nevertheless, we agree that robustness to imperfect annotations is relevant. Following your suggestion, we evaluated two perturbation settings:

Reference annotation settingISICDeepGlobe
Original64.057.6
BBox jitter63.257.3
Mask noise63.756.8

The decreases are limited to 0.3–0.8 mIoU, supporting robustness to the tested perturbations.

Question 2 [Validity of correspondence reliability] Following your suggestion, we analyzed the correlation between correspondence reliability and per-target visual prediction error, defined as 1−IoUvis1-\mathrm{IoU}_{\mathrm{vis}}, using 600 evaluation episodes per dataset in the 1-shot setting.

DatasetPearson rSpearman ρ
ISIC−0.550−0.544
SUIM−0.465−0.440

On these datasets, higher reliability is associated with lower visual prediction error, supporting its use in weighting visual corrections. This analysis validates the reliability score’s predictive association in these settings. We add this analysis in the revised Appendix Sec.A4.4*.

Question 3a [Reference selection] We retain the two references with the highest leave-one-out utility to limit inference cost while prioritizing useful visual evidence. All five references still participate in category adaptation. Following your suggestion, we compared alternative inference strategies:

Reference aggregation strategyISICDeepGlobe
Top-2 by default64.857.4
All references64.957.2
Random two references64.156.6
Similarity-weighted aggregation64.857.5

The default performs comparably to using all references or similarity-weighted aggregation, while random selection reduces mIoU by 0.7 and 0.8 mIoU. These results suggest that performance is not strongly dependent on the exact aggregation strategy; top-2 selection offers a practical accuracy–cost trade-off. We add these discussions to Appendix Sec.A3.6.

Question 3b [Scaling with K] For larger K, reference-averaged adaptation scales approximately linearly per optimization step, while leave-one-out selection evaluates K(K-1) directed reference pairs. This preparation is performed once per reference set. Retaining two references bounds the number of visual comparisons per target at two, whereas using all references requires K.

AuthorsOpenReview ↗

Polite Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your constructive comments and valuable evaluation of our work.

We would like to kindly invite you to take another look at our responses and reconsider your assessment in light of our clarifications. We would greatly appreciate it if you could let us know whether our responses have adequately addressed your concerns, or if there are any remaining questions that we could further clarify.

Thank you again for your time and consideration.

Thank you for the detailed responses and additional experiments, which have addressed my main concerns. I will maintain my original score of 3 (weak acceptance).

Thank you very much for taking the time to read our rebuttal and provide further feedback. We sincerely appreciate your constructive comments and thoughtful consideration of our work. Your feedback has been very helpful in improving our paper.

Reviewer_6xnoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the Cross-Domain Few-Shot Segmentation (CD-FSS) task by proposing a dual-level self-distillation framework (D2D^2-SAM3) based on a frozen SAM3 model. The method uses Category Prototype Distillation (CPD) to transfer GT-box-prompted predictions into a reusable category prototype, and Visual Residual Distillation (VRD) to calibrate reference-target visual correspondences for each category and express them as a residual. Finally, Reliability-aware Fusion (RAF) combines the two predictions based on prediction uncertainty and matching reliability. The effectiveness of the method is validated on four CD-FSS benchmarks and extended tasks.

Critical strengths
  1. Clear problem analysis. Through preliminary experiments, the paper points out that existing visual-matching-based few-shot segmentation methods suffer from large IoU variations across different target appearances and reference selections. It also verifies that category prototypes and pairwise visual matching are complementary across different visual similarity levels, providing a clear motivation.

  2. Small number of trainable parameters and stable performance. With the SAM3 backbone kept frozen, only about 8.2K parameters are optimized per category. Experimental results show that the method achieves high mIoU across four CD-FSS datasets while reducing sensitivity to reference image selection (low variance).

  3. Comprehensive experiments and ablations. The authors evaluate the method on CD-FSS, In-Domain FSS, and Part Segmentation benchmarks, and provide controlled comparisons using the same SAM3 backbone (Table A5) as well as ablation studies for each component.

Critical weaknesses
  1. Limited methodological novelty (Incremental Novelty). The core framework essentially adapts the common "category prototype + visual matching residual + uncertainty/reliability gated fusion" paradigm in few-shot learning to the latest SAM3 model. Although the engineering integration is complete and combined with a self-distillation strategy, individual modules (such as cosine-similarity-based foreground-background matching, quadratic uncertainty weighting, and cycle-consistency checking) are standard combinations of existing techniques and lack core theoretical or methodological breakthroughs.

  2. Not strictly training-free and misaligned baseline comparisons. Although only about 8.2K parameters are optimized per category, the method still requires 80 gradient-based adaptation steps (20 + 20 + 40) for each new category, taking about 2 minutes. In the Vision Foundation Model group of Table 3, many compared baselines (e.g., Matcher, GF-SAM, PerSAM, INSID3) are strictly training-free methods. Outperforming strictly training-free baselines after introducing test-time gradient updates and hyperparameter tuning is expected, but this weakens its convenience as a plug-and-play few-shot segmentation method in practical deployment.

  3. Complex self-distillation and fusion pipeline with many hyperparameters. The reference adaptation and inference pipeline involves multiple multi-stage designs and hyperparameters (such as preliminary prototype warm-up, greedy teacher query selection, 6 deterministic view augmentations, multiple loss weights λd,λp,λr,λv,λf,λg\lambda_d, \lambda_p, \lambda_r, \lambda_v, \lambda_f, \lambda_g, and temperature coefficients). This multi-stage engineering stacking increases the complexity of reproducing and transferring the method to other segmentation foundation models.

Questions for authors
  1. In Table 3, most baselines based on vision foundation models are strictly training-free. How would the performance gap change if other baselines were also equipped with similar lightweight test-time tuning, or if the proposed method were modified into direct feature aggregation without gradient optimization?

  2. The overall framework involves three optimization stages (20 steps for preliminary prototype + 20 steps for CPD + 40 steps for VRD) and numerous hyperparameters. Can the training pipeline be simplified (e.g., end-to-end single-stage joint optimization), and what would be the impact on final performance?

  3. In the 1-shot to 5-shot experiments (Table 3), increasing from 1-shot to 5-shot brings very marginal gains for the proposed method (only ~1% mIoU on average), whereas other methods usually show more noticeable improvements. Does this indicate a bottleneck in the current category prototype distillation mechanism when leveraging diverse information from multiple reference images?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6xno (1/2)

We thank the reviewer for recognizing our clear problem analysis, compact adaptation, stable performance, and comprehensive experiments. We appreciate the constructive feedback and address the concerns below.

Weakness 1 [Methodological novelty] Our contribution has two distinct aspects. First, a new problem insight: we identify a stability–adaptability trade-off in CD-FSS. Pairwise visual matching adapts to target appearance but is sensitive to reference–target differences, whereas a reusable category prototype provides stability but cannot fully capture appearance variations (refer to Figs.1–2 and Tab.2). Second, a method designed to address this trade-off: our dual-level distillation framework learns a reusable category prototype through CPD and category-calibrated visual corrections through VRD. The methodological novelty lies in jointly distilling reusable semantic prompts and category-specific visual calibration for CD-FSS, with visual correspondence serving as a controlled correction to the category prediction. RAF regulates this correction according to correspondence reliability and prototype uncertainty. This contribution lies in the problem-driven formulation and coordinated design that effectively addresses the identified trade-off, supported by component ablations and backbone-controlled comparisons.

Weakness 2 & Question 1a [Not training-free and deployment] We explicitly acknowledge in Appendix Sec.A1 that D2-SAM3 is not training-free. We adopt lightweight reference adaptation to address the limitations of fixed pretrained representations under domain shifts, trading a small preparation cost for improved segmentation performance. Adaptation takes approximately 35 seconds per category (2-min preparation time in Tab.6 covers three ISIC categories), optimizing only 8.2K parameters. With this modest adaptation cost, D2-SAM3 improves over direct SAM3 visual prompting by 10.7 mIoU on average across the four CD-FSS benchmarks (Tab.A7). As clarified in revised Sec.5.6, D2-SAM3 uses fewer total parameters than Matcher and GF-SAM (840M vs. 945M). These measurements support practical use when adaptation is amortized across target images.

Question 1b [Applicability to other backbone models] Our designs follow the identified stability–adaptability trade-off: we need a reusable category representation alongside target-specific visual correspondence. SAM3 provides a text-prompt space in which we can learn this representation while keeping the backbone frozen. Our current formulation depends on this capability and therefore cannot be directly transferred to SAM1 or DINOv3 without redesign.

Question 1c [Training-free alternatives] We explored two training-free alternatives. First, Appendix Sec.A2.2 evaluates category names as alternative text prompts directly with frozen SAM3. Their poor performance indicates that manually specified text alone does not provide an effective category prototype under domain shifts. Second, Appendix Sec.A4.9 evaluates GF-SAM with SAM replaced by SAM3, alongside direct SAM3 visual prompting. These alternatives remain below D2-SAM3, showing that access to SAM3’s stronger pretrained capabilities alone is insufficient to match our results. Although these controls do not exhaust optimization-free feature aggregation strategies, they motivate our lightweight adaptation as a practical balance between rapid deployment and segmentation accuracy.

AuthorsOpenReview ↗

Response to Reviewer 6xno (2/2)

Weakness 3a & Question 2a [Multi-stage pipeline] Our multiple stages have distinct roles: preliminary adaptation establishes the correction and preservation regions, CPD refines the category prototype, and VRD calibrates visual residuals against the fixed prototype. Following your suggestion, we evaluated joint optimization:

Optimization strategyISICDeepGlobe
Joint optimization57.248.3
Sequential optimization (ours)64.057.6

Joint optimization reduces performance by 6.8 and 9.3 mIoU, respectively. This is consistent with interference between objectives: updating the prototype while calibrating its visual correction creates a changing reference prediction, making the objectives harder to balance. Sequential optimization separates these updates while remaining efficient, requiring approximately 35 seconds per category and no gradient updates during inference (refer to revised Sec.5.6).

Weakness 3b & Question 2b [Hyperparameters] We use separate symbols to distinguish the roles of different loss terms, but λp=λf=λg=1.0, and the six loss weights are shared across the four CD-FSS datasets. The six augmented views also form a fixed deterministic set. Appendix Sec.A4.5 and Tab.A5 show that performance is stable across the tested ranges of seven main hyperparameters, with variations of at most 0.9 mIoU within each parameter sweep on each dataset. These results indicate limited sensitivity to the tested settings.

Question 3 [Performance gains from more references] As discussed in Sec.5.3, D2-SAM3 achieves strong 1-shot mIoU, averaging 98.5% of its 5-shot mIoU. We treat it as a strength in annotation efficiency, rather than interpreting the modest 5-shot gains as a weakness. D2-SAM3 learns an effective category prototype and visual residual branch to mitigate reference-target visual fluctuations from just one reference, supporting the effectiveness of prototype distillation under scarce supervision. Our method addresses unreliable reference–target matching through a reusable category prototype and reliability-controlled visual corrections, reducing dependence on finding a visually similar reference. This interpretation is supported by the similarity analysis in Tab.2 and reference-selection robustness in Tab.5. Additional references improve performance on all four datasets. We have added this to the revised Sec.5.3.

AuthorsOpenReview ↗

Polite Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your constructive comments and valuable evaluation of our work.

We would like to kindly invite you to take another look at our responses and reconsider your assessment in light of our clarifications. We would greatly appreciate it if you could let us know whether our responses have adequately addressed your concerns, or if there are any remaining questions that we could further clarify.

Thank you again for your time and consideration.

Reviewer_6W49OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new method for cross-domain few-shot segmentation. It adapts a frozen SAM3 using a learned category prompt, a calibrated visual correspondence branch, and reliability-weighted fusion. To bridge the stability-adaptability trade-off between instance-sensitive pairwise visual matching (which fails when reference and target appearances diverge) and static category prototypes (which provide consistent predictions across targets but fail to capture fine-grained appearance variations) while keeping SAM3 completely frozen, the authors introduce a two-stage reference adaptation and target inference pipeline optimizing only 8,195 parameters per category (32×25632 \times 256 prototype tokens + 33 scalar calibration weights). Experiments across 4 CD-FSS benchmarks (ISIC, Chest X-Ray, DeepGlobe, SUIM), 3 in-domain FSS benchmarks (PASCAL-5i5^i, COCO-20i20^i, FSS-1000), and 2 one-shot part segmentation benchmarks (PASCAL-Part, PACO-Part) show empirical gains and reduced variance across reference selections.

Critical strengths
  1. Clear Empirical Motivation: Fig. 2 isolates reference sensitivity by fixing target images while varying references. Table 2 controls for the SAM3 backbone and shows that prototypes and visual matching excel under different similarity conditions, supporting their combination.

  2. Broad Evaluation and Backbone Controls: The evaluation spans nine benchmarks. Table A5 provides evidence that the gains extend beyond upgrading to SAM3, with substantial improvements over direct visual prompting on DeepGlobe and SUIM.

Critical weaknesses
  1. Contribution of Teacher Supervision: I am concerned that Table A2 does not isolate whether CPD benefits from the teacher’s soft predictions or from selective supervision and prediction preservation. A potentially useful control may be to replace the teacher targets with ground-truth labels on the same pixels, while keeping the selection masks, preservation loss, regularization, and optimization unchanged.

  2. Evaluation Protocol and Runtime Accounting: The method reuses category-specific adaptation across target images, but the comparability of its support/query sampling and metric aggregation with the published baselines is unclear. It would be helpful if the authors could specify these details and clarify whether Table 6’s preparation time is per support set, category, or benchmark, separating adaptation costs from inference.

Questions for authors
  1. The current ablations do not clearly separate the benefit of teacher predictions from selective supervision and prediction preservation. I am wondering does the full method outperform a teacher-free baseline supervised directly by the reference masks, with the architecture and optimization settings held fixed?

  2. Currently it seems that RAF downweights visual corrections when the category prediction is confident, which may limit its ability to recover from confident errors or empty predictions. Could the authors clarify how often does the category branch make confident errors or retain no queries, and how does RAF handle these cases?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 6W49

We thank the reviewer for recognizing our empirical motivation, the complementary strengths of category prototypes and visual matching, and our broad evaluations. We appreciate the constructive suggestions and address the concerns below.

Weakness 1 & Question 1 [Contribution of teacher soft predictions] Thank you for suggesting this control experiment. To isolate the benefit of teacher soft predictions, we replaced them with reference GT labels on the same selected pixels, keeping the selection masks, preservation loss, regularization, architecture, initialization, and optimization settings unchanged. The results are summarized below:

Prototype learning variantISICDeepGlobe
Preliminary Prototype (Tab.A3)58.155.7
Naive Teacher Distillation (Tab.A3)60.956.2
Reference GT mask (new)61.256.6
Full CPD (Tab.A3)64.057.6

Full CPD exceeds this control by 2.8 and 1.0 mIoU, demonstrating the additional benefit of teacher soft predictions under identical selection and preservation settings. This control also improves over the Preliminary Prototype, supporting the benefit of the supervision strategies. We have added these results and analysis to the revised Appendix Sec.A4.3 and Tab.A3, highlighted in red.

Weakness 2a [Evaluation protocol] In the K-shot setting, we randomly select K annotated references per category for distillation and reuse the resulting category-specific parameters across target images. During inference, we use exactly the same reference–target pairs and evaluation metrics as the existing methods (Matcher, GF-SAM). The reference and target sets are disjoint. These consistent sampling and evaluation protocols ensure a fair comparison. We have added these results and analysis to the revised manuscript Sec.5.2, highlighted in red.

Weakness 2b [Runtime] Tab.6 reports runtime on ISIC under the 1-shot setting. The preparation time covers adaptation for all three ISIC categories: distillation takes approximately 35 seconds per category, reported as approximately 2 minutes. This cost is incurred once for the sampled reference set and reused across target images. The subsequent 5 minutes cover inference on the full evaluation set of 600 images and are measured separately from adaptation. We have clarified this in the revised Sec.5.6, with the changes highlighted in red.

Question 2 [Confident errors and empty query selection] When no query passes the confidence threshold, we retain the highest-confidence query rather than constructing an all-zero category prediction. This follows the CD-FSS evaluation setting, where each target image must contain the queried category. We have clarified this fallback in the revised Appendix Sec.A3.4. RAF addresses category-prediction uncertainty by introducing reliable visual corrections, assuming that confident category predictions are generally trustworthy. We acknowledge that confident errors can violate this assumption and may remain uncorrected in some corner cases. Nevertheless, the RAF ablations in Tab.4 and Tab.A3 demonstrate its overall benefit; they do not imply recovery from every failure case.

AuthorsOpenReview ↗

Polite Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your constructive comments and valuable evaluation of our work.

We would like to kindly invite you to take another look at our responses and reconsider your assessment in light of our clarifications. We would greatly appreciate it if you could let us know whether our responses have adequately addressed your concerns, or if there are any remaining questions that we could further clarify.

Thank you again for your time and consideration.

Thank you for addressing my concerns. I will keep my original score.

Thank you very much for taking the time to read our rebuttal and provide further feedback. We sincerely appreciate your constructive comments and thoughtful consideration of our work. Your feedback has been very helpful in improving our paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to Area Chairs and All Reviewers

We thank the Area Chairs and all Reviewers for their careful evaluation and constructive feedback. We appreciate the recognition of our clear problem analysis and motivation (Reviewers 6W49, 6xno, and sqH1), compact adaptation and strong, stable performance (Reviewers 6xno and sqH1), and comprehensive evaluation with component ablations and backbone-controlled comparisons (Reviewers 6W49, 6xno, and sqH1). We also thank Reviewer sqH1 for highlighting the rationale for modeling visual correspondence as a residual correction.

We provide individual, point-by-point responses below to address each reviewer’s concerns, supported by additional experiments and clarifications. We have also revised both the main text and appendix to incorporate relevant suggestions, clarify methodological and evaluation details, and present the new results and analyses. These changes are highlighted in red in the revised manuscript.

Dear Authors and Reviewers (6W49, 6xno, sqH1),

I commend the authors for providing such detailed responses and for their polite follow-up comments. It is clear that a significant amount of effort has been put into addressing the reviewers' concerns.

To Reviewers 6W49, 6xno, and sqH1: The authors have specifically addressed your questions and requested your feedback. As we approach the end of the discussion period, I kindly request that you take a moment to review their responses and leave a brief note on OpenReview.

A simple confirmation stating whether your concerns have been addressed and if your score remains the same is completely sufficient. If I do not receive a follow-up from you before the discussion period closes, I will proceed with my evaluation based on the original reviews and the authors' written responses.

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.