acceptodds
Under review as a conference paper at ICLR 2027

Beyond What to Select: Plug-and-play Oscillatory Data-Volume Scheduling for Efficient Model Training

Abstract

Data selection accelerates training by prioritizing informative training samples, yet existing methods primarily optimize *which samples to select* while keeping *how much data to use* fixed throughout training. As a result, the temporal role of data volume remains largely underexplored, even in dynamic selection. In this work, we find that the instantaneous data volume directly shapes subset-based optimization dynamics, revealing a regularization–fidelity trade-off: lower ratios induce stronger stochastic regularization while reducing computation, whereas higher ratios provide more faithful and stable updates. Therefore, fixing data volume throughout training constrains optimization and can lead to suboptimal performance under a given training budget, motivating a rethink of data selection beyond fixed-ratio training. Building on this insight, we propose **PODS**, a **P**lug-and-play **O**scillatory **D**ata-volume **S**cheduling framework that alternates low- and high-ratio phases while preserving a prescribed cumulative data budget. Given a target ratio, the schedule is analytically determined without dataset-specific tuning and can be directly integrated with existing static or dynamic selectors. Experimental results across diverse vision and language tasks show that **PODS** consistently extends the efficiency frontier of fixed-ratio selection, reducing ImageNet-1k training cost by 40% while improving accuracy by 0.4%, and achieving approximately 2× faster LLM instruction tuning without performance degradation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dynamic Data Selection, Data Selection, Dataset Pruning

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

This paper proposes PODS, a dynamic data-volume scheduling method that periodically alternates between low- and high-ratio training phases to improve training efficiency under a prescribed cumulative data budget. The authors explain the strategy through the trade-off between gradient stochasticity and optimization fidelity and derive an analytical scheduling rule. Experiments across image classification, object detection, and LLM fine-tuning demonstrate that PODS can preserve or improve model performance while reducing training costs.

Critical strengths
  1. The method is simple and practically useful, PODS requires no additional scheduling parameter search and can be integrated with existing data selection methods. Its simple design makes it easy to incorporate into existing training pipelines.
  2. Extensive experimental evaluation. The paper evaluates the effectiveness of PODS across diverse tasks and model architectures.
Critical weaknesses
  1. Prior studies, such as SCAN (ICCV 2025) [1] and Music (Pattern Recognition 2024) [2], have already explored dynamically adjusting the data selection ratio during training. Although the two-phase oscillatory schedule in PODS has some distinctive characteristics, the paper does not sufficiently discuss its differences from these existing approaches.

  2. Proposition 3.1 derives the effect of the selection ratio on gradient covariance based on the average gradient of a uniformly sampled subset. However, the actual method employs mini-batch optimization with a fixed batch size and loss-based hard mining. Under uniform sampling, the marginal covariance of an individual mini-batch gradient does not directly depend on the size of the intermediate subset. Therefore, the current theoretical analysis does not sufficiently explain how oscillatory data-volume scheduling induces the claimed stochastic regularization effect in the actual training procedure, nor does it establish a direct connection between this effect and improved generalization.

[1] Guo, Y., & Kankanhalli, M. (2025). SCAN: Bootstrapping Contrastive Pre-training for Data Efficiency. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 3662–3672.

[2] Yin, Z., Pu, J., Wan, R., & Xue, X. (2024). Embrace Sustainable AI: Dynamic Data Subset Selection for Image Classification. Pattern Recognition, 151, 110392.

Questions for authors
  1. What is the core novelty of PODS compared with existing dynamic selection-ratio scheduling methods such as SCAN? What are the advantages of the specific oscillatory schedule?
  2. How does Proposition 3.1 relate to the actual mini-batch optimization procedure and loss-based hard mining? Can the authors further explain the mechanism by which the selection ratio affects gradient stochasticity and model generalization in practice?
4 comments

Dear Reviewer NtKB,

Thank you for the careful review and constructive suggestions. We appreciate your recognition of our work's strengths, e.g., practical effectiveness, simple and tuning-free design, seamless integration with existing training pipelines, and extensive empirical validation.

We provide responses to address the comments below:

  • Q1/2: Differences from SCAN [a] and MUSIC [b].

  • A1/2: Thanks for suggesting additional related works. We would like to clarify that while SCAN employs iterative dataset mutation with cosine-based pruning for contrastive pre-training, and MUSIC adaptively determines subset ratios through performance-driven feedback for image classification, PODS explicitly investigates oscillatory data-volume allocation under a fixed cumulative training budget.

    Our key advantage lies in its principled, lightweight, and broadly applicable scheduling design: PODS periodically alternates between selective training and broader data coverage, without auxiliary optimization, feedback-driven ratio adjustment, or dataset-specific tuning. Moreover, PODS decouples data-volume scheduling from sample importance estimation, enabling direct integration with existing advanced selectors. Experiments in Sec. 4 further demonstrate its broad applicability across vision and language tasks. Thus, PODS advances dynamic data selection through a simple, budget-constrained oscillatory scheduling framework with independently validated effectiveness and broad applicability.

Following your suggestion, we have included both works [a,b] in Section 2 of the revised manuscript. Specifically,

  • Sec 2, paragraph 2, line 142-143: add reference [a] "iterative dataset mutation with cosine-based ratio scheduling for contrastive pre-training Guo & Kankanhalli (2024),"
  • Sec 2, paragraph 2, line 144: add reference [b] "performance- guided adaptation of selection ratios Yin et al. (2024)"
  • Q3: Connection between theoretical analysis and practical mini-batch optimization.
  • A3: Insightful comment. We would like to clarify that Prop. 3.1 characterizes stage-wise subset-gradient variability, rather than the marginal variance of individual mini-batch gradients, which need not depend on the intermediate subset size under uniform sampling with a fixed batch size. In practical hard mining, however, varying the selection ratio changes the selected data distribution, thereby affecting the expected gradient direction, within-subset variability, and data coverage during mini-batch optimization.

Our experiments provide direct empirical support. As shown in Fig. 5, measured gradient variance during low-ratio phases is 2.0×/2.6× higher than during high-ratio phases. Importantly, suppressing this variability consistently degrades performance, supporting its contribution to the observed performance gains. Moreover, Table C-1 below demonstrates that oscillatory scheduling improves performance even with random selection, establishing its effectiveness independently of hard mining.

We have clarified the theoretical scope in Sec. 3.2 of the revised manuscript. Together, these results connect stage-wise data-volume control to practical optimization dynamics, with controlled experiments demonstrating the effectiveness of oscillatory scheduling.

Table C-1: Effectiveness of PODS with random selection strategies.

MethodTiny-IN1kC-100ImageNet-1k
30%50%70%30%50%70%60%
Full Dataset45.078.276.4
PODS + Random42.045.947.878.279.179.576.8
PODS + Hard Mining44.546.949.078.479.179.476.8
AuthorsOpenReview ↗

Looking forward to the reply

Dear Reviewer NtKB,

Thank you again for your time and constructive feedback. Following your suggestions, we have conducted additional experiments and further discussed related work. We have also revised Sec. 2 of the manuscript to better highlight the distinctions of PODS.

As the discussion period is approaching its end, we would greatly appreciate your feedback on whether our responses have adequately addressed your concerns. If there are further concerns or questions, we are willing to address them.

Thanks again for taking the time to review our work and provide insightful comments.

Authors of Submission 3402

Thanks for the detailed response. My concerns have been mostly addressed, and I've decided to change the rating to weak accept.

Dear Reviewer NtKB,

Thank you very much for your positive feedback and recognition of our work. We are pleased that our responses and revisions have addressed your concerns. We sincerely appreciate your constructive suggestions, which have helped strengthen our work.

Best regards,

Authors of Submission 3402

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

The paper proposes PODS, a scheduling layer that treats the per-epoch selection ratio p_t as a time-varying control variable. Under a fixed cumulative budget, it alternates low-ratio epochs (p_low) with a high-ratio refinement epoch (p_high = 1−ε, period k*+1), with k* and p_low obtained in closed form. A one-step Taylor expansion under uniform subsampling yields a ratio-dependent term λ(p)=(1−p)/p multiplying Tr(HC), which the authors interpret as a regularization–fidelity trade-off. Sample selection is decoupled (default: top-p_t loss), so PODS can be stacked on existing selectors. Experiments span 21 benchmarks including ImageNet-1k, long-tailed and OOD recognition, COCO detection, and LLM SFT/reasoning fine-tuning, and report accuracy parity or improvement at 30–90% budgets with 40% training-cost reduction on ImageNet-1k.

Critical strengths

(1) The idea is simple, well motivated, and practically attractive: the schedule is analytic, tuning-free given p and ε, and the budget constraint is exactly satisfied per period. (2) Experimental breadth is exceptional — vision and language, ResNet/ViT/LLM, balanced/long-tailed/OOD, classification/detection/reasoning — and both accuracy and wall-clock/GPU-hour costs are reported. (4) The paper is transparent about overheads and about failure at very low ratios (Appendix L).

Critical weaknesses

(1) The theory does not cover the algorithm: Prop. 3.1 assumes uniform sampling without replacement, while the default selector is biased top-p_t loss mining, whose bias also changes with p_t, so "data volume" and "subset composition" are confounded. The interpretation of the non-negative term R ≥ 0 as beneficial regularization is not justified — it is an excess-loss term, and Cor. B.1 only shows non-negativity/monotonicity. (2) No variance estimates anywhere; the headline claims of beating full-data training (ImageNet 76.8 vs 76.4, long-tailed, fine-grained, Tiny-ImageNet 47.8 vs 45.0 for oscillation-only) are within or below typical run-to-run noise, and some baseline numbers are quoted from other papers rather than re-run. (3) No comparison against alternative schedule shapes (monotone, smooth, random) at the same budget, so "oscillation is the mechanism" is not established; the ε-sensitivity in Appendix J arguably undercuts it. (4) The extreme schedule (p_low = 5% at p = 0.5) and the 5.4-second overhead claim are not reconciled with what the method must compute.

Questions for authors

The scheduling idea is simple and useful, but the theoretical framing does not support it: Prop. 3.1 assumes uniform subsampling while the method performs biased top-p_t mining, and the one-step expansion is a textbook variance identity whose (1−p)/p term already appears in InfoBatch's Appendix B.4, where it is treated as a cost rather than a benefit. The evidence is equally fragile: every result is a single run, several baselines are quoted from other papers rather than re-run, no monotone, smooth or random schedule is compared at equal budget, OrderDP — the state of the art at 50–70% pruning — is absent, and Table 5 reports three of four LLM metrics identical to full-data training at half the steps. Without seeds, a schedule-shape ablation, an OrderDP comparison and a "fixed ratio + terminal annealing" control, the claim that oscillation pushes the efficiency frontier is not established. I consider to raise my score if concern can be fixed.

6 comments

Dear Reviewer wopD,

We sincerely thank you for the careful review and insightful suggestions. We appreciate your recognition of our work's strengths, e.g., well-motivated method, particularly attractive design, and exceptional experimental breadth.

We provide responses to address the comments below:

  • Q1: Theoretical scope and applicability to hard mining.
  • A1: Thanks for the comment. We would like to clarify that uniform sampling is adopted to isolate the stage-wise effect of data volume from sample-selection bias, rather than to characterize arbitrary selection strategies. Accordingly, we have revised Sec. 3.2 on page 4 to clarify the theoretical scope and characterize the variance-related term as a local stochastic correction.

Importantly, we emphasize that PODS is a selector-agnostic framework whose effectiveness does not depend on hard mining. As shown in Table B-1 below, even with random selection, PODS consistently outperforms full-data training across all ratios. These controlled comparisons directly establish the independent effectiveness of oscillatory scheduling. Furthermore, as we have shown in Table 7 in Sec. 4.3, PODS can be integrated with existing more advanced static and dynamic selectors and consistently enhance their performance, confirming its broad applicability and plug-and-play compatibility.

Together, these results establish temporal data-volume scheduling as an effective, generalizable, and plug-and-play optimization dimension beyond sample importance estimation.

Table B-1: Effectiveness of PODS with random selection.

MethodTiny-IN1kC-100ImageNet-1k
30%50%70%30%50%70%60%
Full Dataset45.078.276.4
PODS + Random42.045.947.878.279.179.576.8
PODS + Hard Mining44.546.949.078.479.179.476.8
  • Q2: More results on variability.

  • A2: Thanks for pointing this out. We would like to clarify that our results are obtained via 5 independent runs on CIFAR-10/100 and 3 runs on the other evaluated settings, and we report mean±std in Table B-2 below. The results demonstrate consistent performance with low run-to-run variability. The standard deviations are consistently small (0.1–0.3 points) relative to the principal gains, including +1.2 points on CIFAR-100 and +4.0 points on Tiny-ImageNet at the 70% ratio. For LLM fine-tuning, PODS matches full-data performance with a std of only 0.1-0.2, while reducing training costs by nearly 50%.

    These results confirm that the reported improvements are stable and robust to training variability.

Table B-2: Mean accuracy (%) ± standard deviation across independent runs.

Dataset30%50%70%Full Dataset
CIFAR-1095.1±0.295.4±0.195.9±0.195.6±0.1
CIFAR-10078.4±0.279.1±0.179.4±0.178.2±0.2
Tiny-ImageNet44.5±0.346.9±0.249.0±0.245.0±0.2
60%Full Dataset
ImageNet-1k76.8±0.276.4±0.2
50%Full Dataset
Qwen2.5-7B74.0±0.274.0±0.2
LLaMA-3.2-3B55.9±0.156.0±0.1
  • Q3: Further comparison with different scheduling strategies.
  • A3: Thanks for your suggestion. As suggested, we conducted additional matched-budget comparisons with random, monotone, and smooth scheduling strategies (Table B-3).
    It can be seen that PODS consistently outperforms all alternatives across all three selection ratios. These results demonstrate that the advantage stems from the specific oscillatory design rather than merely varying data volume. Furthermore, the ε-sensitivity results in Appendix J confirm robust performance across different oscillation margins, further supporting the effectiveness of our scheduling strategy.

Table B-3: Comparison of alternative scheduling strategies at 30% and 50% on CIFAR-100.

Sel. RatioRandom ScheduleMonotone ScheduleSmooth ScheduleOurs (PODS)
30%76.976.576.578.4
50%78.378.577.679.1
70%78.778.778.779.1
  • Q4: Clarification on the overhead.

  • A4: We would like to clarify that the reported 5.4s overhead refers to the additional scheduling and selection cost, rather than the overall training cost, following the overhead accounting convention of existing methods such as InfoBatch. Importantly, PODS analytically determines the entire ratio schedule in advance and requires no auxiliary optimization or additional forward passes, as sample scores are updated during standard training. Thus, extreme low/high-ratio alternation only changes data exposure, without introducing additional scoring computation. Thus, the scheduling overhead remains negligible.

  • Q5: Comparison and integration with OrderDP [a].

  • A5: Thanks for suggesting more related work. We have conducted additional evaluations covering both direct comparison and integration with OrderDP.

1. As shown in Table B-4 below, PODS consistently outperforms OrderDP across all selection ratios, achieving 1.2-1.7% on C-100 and 0.4% on ImageNet-1K, demonstrating its competitive accuracy-efficiency performance.

2. More importantly, PODS can further enhance OrderDP rather than merely compete with it. As shown in Table B-5, integrating PODS with OrderDP consistently improves its performance across all selection ratios. These results further establish temporal data-volume scheduling as an effective and complementary optimization dimension.

3. As suggested, we have included the suggested work [a] in Sec. 2 and Table 1 of the revised manuscript. Specifically,

  • Sec 2, paragraph 2, line 140-141: add reference [a] "OrderDP Jin et al. (2026) combines random and importance-based pruning to mitigate selection bias"

We have comprehensively incorporated OrderDP into our experimental evaluations in Sec. 4 of the revised manuscript, including direct comparisons and integration studies in Tables 1, 2, and 7, further strengthening the empirical validation of PODS against this strong baseline.

Table B-4: Accuracy (%) comparison with OrderDP. We use the reported accuracy from [a].

MethodCIFAR-10CIFAR-100ImageNet-1K
Selection Ratio30%50%70%30%50%70%60%
OrderDP95.095.395.676.777.978.276.4
PODS (Ours)95.195.495.978.479.179.476.8
Improvement+0.1+0.1+0.3+1.7+1.2+1.2+0.4

Table B-5: Accuracy (%) comparison with OrderDP.

Selection Ratio30%50%70%
OrderDP76.777.978.2
PODS+OrderDP78.279.279.4
Improvement+1.5+1.3+1.2
AuthorsOpenReview ↗

Looking forward to the reply

Dear Reviewer wopD,

Thank you again for your time and constructive feedback. In response to your comments, we have conducted additional experiments on training variability, alternative scheduling strategies, and OrderDP comparisons, and clarified the theoretical concerns. We have also incorporated these updates into the revised manuscript.

As the discussion period approaches its end, we would greatly appreciate your feedback on whether our responses have addressed your concerns. We would be happy to address any remaining questions.

Thanks again for taking the time to review our work and provide insightful comments.

Authors of Submission 3402

Reviewer_z6bzOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

In this paper, the authors claim that using dynamic data volume during training can improve the training efficiency. The authors try to prove that dynamic data volume has a regularization effect on training. They conduct a series of experiments to compare the training efficiency with and without dynamic data volume strategy.

Critical strengths
  1. This paper is well-written and easy to read.
  2. The empirical results that using dynamic data volume can improve training efficiency is interesting.
  3. The proposed method can integrated with various data selection methods flexibly.
Critical weaknesses
  1. There exists a significant error in the theoretical result, i.e., Prop 3.1. I need to point out that Prop 3.1 only shows that using dynamic data volume can lower the speed of the expected training loss decreasing. This result is already known, essentially the same results can be found in various analysis on SGD, i.e., the variance of mini-batch can reduce the speed of the expected loss decreasing in each iteration. Lots of advanced SGD, such as SVRG, try to reduce this variance to accelerate training and therefore this is a negative effect. However, the authors have a significant error on this result, that is they view it as a useful regularization effect.
  2. The authors need to explore whether advanced dynamic data volume strategy can further improve the training efficiency.
  3. In the theoretical analysis, the authors assume that the training data are uniformly sampled during training, which is inconsistent with the settings in real application of data selection. The authors are recommended to analyze their method when integrated with advanced data selection algorithms.
Questions for authors

refer to weaknesses.

4 comments

Dear Reviewer z6bz,

We sincerely thank you for the careful review and insightful suggestions. We appreciate your recognition of our work's strengths, e.g., practical significance, flexibility and broad compatibility, and clarity of presentation.

We provide responses to address the comments as follows:

  • Q1: Theoretical interpretation of Prop. 3.1.

  • A1: Thanks for your careful review. We have clarified that the variance-related term characterizes a local stochastic correction in the revised manuscript.

    Importantly, the effectiveness of oscillatory scheduling is independently established through controlled experiments, rather than inferred from this theoretical correction. As shown in Table A-1, even with random selection, PODS consistently outperforms fixed-ratio selection across all three budgets (30%/50%/70%), achieving gains of +0.5/+3.1/+4.7% on Tiny-ImageNet and +3.5/+3.8/+2.2% on CIFAR-100, respectively. These results directly demonstrate that the benefit of oscillatory scheduling does not depend on hard mining. Furthermore, Table 7 demonstrates that PODS consistently improves existing static and dynamic selection methods across various ratios, demonstrating its broad compatibility with advanced selection strategies.

    Together, these results establish temporal data-volume scheduling as an effective and complementary optimization dimension, whose empirical benefits are independently validated rather than assumed from the theoretical analysis.

Table A-1: Effectiveness of PODS with random selection strategies.

MethodTiny-IN1kC-100ImageNet-1k
30%50%70%30%50%70%60%
Full Dataset45.078.276.4
PODS + Random42.045.947.878.279.179.576.8
PODS + Hard Mining44.546.949.078.479.179.476.8
  • Q2: More advanced dynamic data volume strategy.

  • A2: Thanks for the suggestion. As we have discussed in Appendix L "Discussion and Future Work", PODS establishes a simple, effective, and practical foundation for dynamic data-volume scheduling, with analytically determined schedules, negligible overhead, and broad compatibility with existing selectors. Exploring more sophisticated adaptive scheduling strategies, beyond the scope of this work, represents a promising direction for future research.

  • Q3: Additional results with advanced data selection algorithms.

  • A3: Thanks for the comment. We would like to clarify that PODS has already been integrated with various advanced static and dynamic selection methods, including Moderate, Glister, UCB, ε-Greedy, and InfoBatch, as reported in Table 7 (Sec. 4.3). PODS consistently improves all these methods across selection ratios, with gains of up to 7%, demonstrating its broad applicability and effectiveness beyond uniform sampling and the default hard-mining strategy.

AuthorsOpenReview ↗

Looking forward to the reply

Dear Reviewer z6bz,

Thank you again for your careful review and constructive feedback. In response to your comments, we have provided additional experimental analyses and clarified the theoretical concerns in our rebuttal and revised manuscript.

As the discussion period approaches its end, we would greatly appreciate your feedback on whether our responses have addressed your concerns. We would be happy to clarify any remaining questions.

Thanks again for taking the time to review our work and provide insightful comments.

Authors of Submission 3402

AuthorsOpenReview ↗

Summary of Additional Evidence and Revisions

Dear Reviewer z6bz,

Thank you again for your constructive feedback and recognition of our work. Considering the limited time available before the discussion period closes and to save your time, we briefly summarize how we have addressed your comments through additional experiments, theoretical clarifications, and manuscript revisions.

  1. Theoretical clarification: We have refined Sec. 3.2 and Appendix B to clarify the scope of Prop. 3.1 and the interpretation of its ratio-dependent stochastic correction.

  2. Independent effectiveness of scheduling: We provide controlled experiments demonstrating that PODS remains effective even with random selection, surpassing full-data training at a 50% budget on both Tiny-ImageNet and CIFAR-100 (Table A-1).

  3. Existing integration with advanced selectors: We would like to highlight that PODS has already been integrated with various advanced static and dynamic selection methods in the original submission (Table 7, Sec. 4.3), consistently improving their performance with gains of up to 7.0%.

These additional results and revisions directly address the key concerns raised in your review and further strengthen the theoretical clarity, empirical evidence, and practical significance of PODS.

As the discussion period approaches its end, we would greatly appreciate your feedback on whether our responses and revisions have adequately addressed your concerns. We would be happy to clarify any remaining questions.

Thank you again for your time and consideration.

Authors of Submission 3402

AuthorsOpenReview ↗

A sincere request to read our rebuttal

Dear Reviewer,

We are extremely sorry to bother you. As the discussion period is ending in a few hours, we humbly request that you read our rebuttal and the updated paper. We have addressed all your comments. Please let us know if you have any further concerns. If you find our responses satisfactory, we would be grateful if you would consider revising the score.

Thank you!

Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response

General Description:

Dear Area Chairs and Reviewers,

We sincerely thank you for the time and effort in reviewing our work. We greatly appreciate the constructive feedback and the recognition from all reviewers—z6bz (R1), wopD (R2), and NtKB (R3)—highlighting key strengths of our work, including (1) the simplicity, practical utility, and plug-and-play compatibility of PODS (R1, R2, R3), (2) promising empirical effectiveness and extensive validation across diverse tasks and architectures (R1, R2, R3), and (3) the clarity of presentation (R1, R3).

In addition, the concerns are mainly concentrated on (1) theoretical interpretation and its connection to practical optimization (R1, R2, R3), and (2) more experiments and analysis (R2, R3). We have addressed these concerns through targeted manuscript revisions and additional experiments, as summarized below.

Additional Experiments and Analyses:

In the responses, we provide additional experimental results and analyses, including:

  1. More analyses of the oscillatory scheduling mechanism (R1: Q1, R2: Q1, R3: Q3) (Table A/B/C-1)
  2. More analyses on the scheduling shape (R2: Q3)(Table B-3)
  3. Variability ayalysis (R2: Q2)(Table B-2)
  4. Comparison and integration with OrderDP (R2: Q5)(Table B-4/5)

We also clarified the theoretical scope, practical optimization mechanism, and distinctions from related work in the revised manuscript.

Thank you again for your thoughtful feedback and for helping us refine our work further.

Sincerely,

Authors of Submission 3402

Dear reviewers,

The authors have submitted their response. Please carefully check whether the new content has addressed your concerns, particularly the gap between the theoretical analysis and the proposed approach, the numerical comparison against other methods, and the advantages of the proposed scheduling approach. If not, please follow up with the authors ASAP so that the authors can further respond before the discussion period ends (Nov 18).

Best,

AC

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Area Chair and Reviewers,

We sincerely thank you for your time and constructive feedback. We have carefully addressed the reviewers' comments through additional experiments, detailed clarifications, and corresponding revisions to the manuscript.

As the discussion period approaches its end, we would greatly appreciate any further feedback on whether our responses have adequately addressed your concerns. We would be happy to provide additional clarifications or evidence before the discussion closes.

Thank you again for your time and consideration.

Authors of Submission 3402

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.