acceptodds
Under review as a conference paper at ICLR 2027

A Dynamic Example Selection Framework for Compute-Efficient Machine Unlearning

Abstract

Machine unlearning aims to remove designated training data's influence while preserving utility at a lower cost than retraining. In sample-level unlearning, forget and retain examples can share classes and features, making their selection important for balancing forgetting and preservation. Repeated full-data updates incur substantial cost. We ask which examples should receive computation to achieve an existing unlearning method's performance at a lower cost. We observe that the initial low-confidence forget subset is enriched for examples with low true-label confidence after retraining on the retain set. Motivated by this observation, we introduce Dynamic Source Selection (DSS), which refreshes retain and forget subsets using the current model's true-label confidence while preserving the base loss and update rule. DSS+ reduces rescoring by keeping selected examples in training. A Gaussian logistic analysis identifies conditions under which updates on these sources improve prediction agreement with the retrained model while preserving retain performance. Empirically, dynamic selection allocates more training to these sources. Experiments on CIFAR-10 and CIFAR-100 demonstrate that DSS can attain accuracy comparable to full-data updates at lower computational cost across multiple methods, with TOFU experiments extending applicability to language models. These results support dynamic source selection as an effective strategy for efficient unlearning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, efficient machine learning, data selection, data privacy

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ViBBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a confidence-based sample selection framework to accelerate existing machine unlearning methods. It dynamically selects retain and forget examples while preserving the underlying unlearning objectives. The authors provide a motivation and evaluate the framework on vision classication and language models, reducing computational time in several settings.

Critical strengths
  1. The framework integrates sample selection into several existing unlearning methods.

  2. Several configurations achieve time savings while maintaining similar model utility.

  3. The evaluation includes computation trajectories and prediction-distribution diagnostics.

Critical weaknesses
  1. Key relevant papers were neither discussed nor compared. RUM [1] groups forget examples by memorization for sequential unlearning, while SULI [2] uses predictive entropy-ranked to order forget subsets for iterative soft relabeling. Neither work is discussed or compared, leaving it unclear what DSS’s dynamic confidence-based selection adds beyond these existing sample scheduling strategies. Bad Teacher [3] has already been updated using a combination of forget data and a subset of retain data.

  2. The theoretical comparison relies on retrained-model confidence, whereas DSS selects samples using current confidence during unlearning. The fixed-source results do not establish why this changing score identifies useful subsequent updates in the actual dynamic procedure.

  3. The MIA results and setup lack transparency, key implementation details of the MIA have not been disclosed, and the results cannot be used to substantiate the arguments. And, the privacy leakage risks associated with excluding certain forget set from the update process were not analyzed.

[1] RUM: What makes unlearning hard and what to do about it (NeurIPS 2024)

[2] SULI: Fine-Grained and Efficient Self-Unlearning with Layered Iteration (IJCAI 2025)

[3] Bad TeacherCan Bad Teaching Induce Forgetting? Unlearning in Deep Networks using an Incompetent Teacher (AAAI 2023)

Questions for authors
  1. What does DSS’s dynamic selection add beyond RUM’s memorization-based grouping and SULI’s uncertainty-based staged unlearning?

  2. What evidence validates current confidence for sample selection during unlearning, given that RUM’s confidence proxy is averaged over the original training trajectory?

  3. Does reducing the retain computation suffice to achieve similar gains, and what exactly is gained by the additional step of filtering the forget data?

2 comments
AuthorsOpenReview ↗

Response to Reviewer ViBB

Thank you for your helpful comments. We have revised the manuscript accordingly and address your questions below.

Q1. RUM and SULI organize unlearning through example groups or stages. DSS instead recomputes selection from the current unlearning model and updates participation in both the retain and forget sets. Examples can enter or leave the selected set, while the base method retains its loss formulation, targets and update rules.

RUM groups forget examples for sequential unlearning. SULI combines uncertainty-based ordering with its own soft-label adjustment procedure. DSS separates the selection policy from the base unlearning method and evaluates that policy across six methods. DSS+ further changes the selection policy by keeping admitted examples and rescoring only those not yet selected. The distinction concerns which pools are selected, when selection is updated and how it is coupled to the underlying unlearning procedure.

Section 2 (Data selection for unlearning) now discusses RUM and SULI alongside DSS. Sections 3.2 and 3.3 specify DSS's refresh procedure and DSS+'s persistent selection, respectively.

Q2. We chose current true-label confidence primarily for its simplicity. It is directly available from a model forward pass and does not require storing or averaging scores over the original training trajectory. It also reflects the model that is about to perform the next unlearning updates.

EL2N and GraNd can also be used as scores in the same selection procedure, as evaluated in Figure 4. Confidence and EL2N use closely related information and produce similar results. Confidence therefore serves as a simple default within the selection procedure.

Section 3.2 now states that true-label confidence is obtained from a forward pass without storing or averaging the original training trajectory. Alternative scoring rules are described in Appendix A.3 and compared in Section 5.2, Figure 4.

Q3. We compared FULL, retain-only selection, forget-only selection and both-pool selection for NegGrad+ on CIFAR-10 and CIFAR-100 with ResNet-18, using ten paired deletion and unlearning requests. The initial model, deletion request, learning rate and base training recipe are shared across policies within each request. The unfiltered pool remains complete.

On CIFAR-100, retain-only and both-pool selection use nearly the same computation, 1.856 and 1.852 PFLOPs, and select similar numbers of retain examples. Their Avg. Gaps are 16.078 and 9.999 pp, respectively. Thus, filtering the forget pool changes the outcome beyond the savings from reducing retain computation alone. Both-pool selection also gives a lower Gap than either one-pool policy on CIFAR-10.

Table R3. Selection-side comparison for NegGrad+ with ResNet-18 under random forgetting. Means over ten paired deletion and unlearning requests. The values are taken from main-text Table 4. Mean selected (R, F) gives the average numbers of available retain and forget examples per epoch.

DatasetSelectionAvg. GapPFLOPsMean selected (R, F)
CIFAR-10FULL1.777 ± 0.0882.969 ± 0.000(45000, 5000)
CIFAR-10Retain only6.261 ± 3.3261.221 ± 0.100(10094.6, 5000)
CIFAR-10Forget only38.039 ± 9.1143.030 ± 0.033(45000, 3710.8)
CIFAR-10Both (DSS)2.129 ± 0.4310.988 ± 0.046(6834.9, 839.0)
CIFAR-100FULL5.933 ± 0.4752.969 ± 0.000(45000, 5000)
CIFAR-100Retain only16.078 ± 0.4021.856 ± 0.008(20406.0, 5000)
CIFAR-100Forget only18.524 ± 1.7243.054 ± 0.007(45000, 3234.2)
CIFAR-100Both (DSS)9.999 ± 1.2321.852 ± 0.031(19866.0, 2342.7)

The expanded selection-side comparison has been added to Section 5.2 (Which set to select), Table 4, including the individual metrics, computation and mean selected counts. The result descriptions explain why reducing retain computation alone does not reproduce the both-pool outcome in these comparisons.

Thank you for the revisions and additional experiments. However, my main concerns remain unresolved.

The paper still lacks comparisons with the newly cited methods, zero-shot and few-shot unlearning methods. The motivation and claimed advantages remain unconvincing.

The new ablation in Table 4 only evaluates NegGrad+ and does not support the broader claims across different unlearning methods.

The response to Q2 does not address my concern about the reliability of current confidence, especially given the findings of RUM. The result of table 26 further raises doubts about this selection rule, and the theoretical analysis still does not justify its effectiveness.

Still failed to change a stronger MIA.

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

To cut the computational cost of machine unlearning, this paper presents Dynamic Source Selection (DSS), which dynamically filters training samples by their true-label confidence. An enhanced variant, DSS+, further reduces scoring overhead by retaining previously selected examples across iterations. Evaluated on CIFAR-10, CIFAR-100, and the TOFU benchmark, DSS integrates effectively with existing unlearning baselines and is supported by a simplified Gaussian logistic regression analysis.

Critical strengths

The proposed framework operates at the data-selection level rather than requiring a new unlearning objective. DSS and DSS+ offer a practical trade-off between adaptive selection and re-scoring overhead. Covering two CIFAR datasets and two network architectures, the vision experiments evaluate six unlearning methods (NegGrad+, NPO, UGradSL+, random labeling, SalUn, and AMUN) across both random and class-concentrated deletion scenarios.

Critical weaknesses

(1) The paper discusses related approaches such as ConMU, SGA, EL2N, GraNd, and InfoBatch. However, the distinction between DSS and existing dynamic selection methods remains insufficiently compelling as a methodological contribution. In particular, the ablation of the selection rule is limited to CIFAR-10 with ResNet-18 across two base methods (NegGrad+ and NPO). Furthermore, Figure 5 fails to demonstrate a clear improvement over EL2N in terms of Avg. Gap and cumulative PFLOPs, and individual evaluation metrics are omitted for this experiment. (2) The Gaussian logistic regression analysis in Section 4.1 is considerably narrower in scope than the proposed general framework; consequently, it cannot serve as a broad justification for low-confidence selection across different unlearning algorithms. (3) While the results demonstrate runtime reductions across several settings, the quality trade-off is not consistently minimal. Furthermore, all unlearning methods evaluated in this work require access to the forget set; as a result, the empirical benefit of integrating DSS/DSS+ with methods that rely solely on the retain set remains unclear. (4) The LLM results establish that the selection procedure can be integrated into language-model unlearning, but they do not establish that it is generally compute-efficient for LLMs. (5) DSS relies on a threshold calibrated from an initial selection fraction ρ\rho, while the refresh interval introduces an additional hyperparameter. Because tuning these parameters depends heavily on the dataset, base method, and training duration, the lack of a principled selection procedure limits ease of use. Given that the framework is presented as method-agnostic, providing a more robust default or a practical calibration heuristic would significantly strengthen the work.

Questions for authors

Please see the weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer yPhT

Thank you for your helpful comments. We have revised the manuscript accordingly and address your questions below.

W1. The base method determines how examples are updated. DSS controls which retain and forget examples participate as the model changes. We treat the unlearning objective and the examples used to compute it as separate design choices. DSS retains the base loss formulation, targets and update rules, while selection changes the empirical objective and actual gradients.

In Figure 4, EL2N and GraNd are used as alternative scores within DSS's selection and refresh procedure. They are not evaluated as separate unlearning algorithms. Confidence and EL2N use closely related prediction information, so similar results are consistent with their roles in this comparison. Our contribution concerns controlling participation across existing unlearning methods, rather than introducing a new scoring function. Confidence provides a simple implementation using the true-label probability from a forward pass.

We have extended the selection-rule comparison to UGradSL+, RL, SalUn and AMUN, covering all six base methods on CIFAR-10 with ResNet-18 under random forgetting. Table R1 below reports the mean endpoint Avg. Gap and cumulative computation over the same ten paired deletion and unlearning requests for all methods. The complete computation–gap trajectories are included in Figure 11 of the revised manuscript (Appendix B.2). This evaluates the selection procedure and alternative scores across a broader range of base update rules. Confidence and EL2N remain broadly similar in the expanded comparison, with differences depending on the base method.

We also compared NegGrad+ and RL on CIFAR-100 with VGG16 over ten paired requests to examine selection under a different dataset and architecture. The endpoint results are shown below in Table R2, with the complete trajectories in Figure 12 of the revised manuscript (Appendix B.2).

Table R1. CIFAR-10 / ResNet-18, random forgetting. Mean ± sample SD across ten paired requests at the reported endpoint. Avg. Gap is in percentage points. The complete trajectories appear in Figure 11 of the revised manuscript (Appendix B.2).

MethodSelectionAvg. GapPFLOPs
NegGrad+FULL1.873 ± 0.1542.969 ± 0.000
NegGrad+DSS1.963 ± 0.1680.992 ± 0.030
NegGrad+DSS+2.086 ± 0.2301.121 ± 0.048
NegGrad+EL2N1.985 ± 0.1770.994 ± 0.038
NegGrad+GraNd2.058 ± 0.2822.053 ± 0.031
NegGrad+InfoBatch1.786 ± 0.2471.983 ± 0.008
NegGrad+Random2.699 ± 0.3021.537 ± 0.006
NegGrad+Fixed initial scores1.909 ± 0.2491.594 ± 0.005
NPOFULL1.652 ± 0.1083.459 ± 0.000
NPODSS1.988 ± 0.1481.070 ± 0.040
NPODSS+1.879 ± 0.2181.215 ± 0.059
NPOEL2N2.051 ± 0.2281.063 ± 0.045
NPOGraNd1.879 ± 0.2062.136 ± 0.059
NPOInfoBatch1.565 ± 0.1122.308 ± 0.005
NPORandom2.730 ± 0.3021.787 ± 0.009
NPOFixed initial scores1.834 ± 0.3011.844 ± 0.007
UGradSL+FULL1.409 ± 0.4752.999 ± 0.000
UGradSL+DSS1.055 ± 0.2692.386 ± 0.052
UGradSL+DSS+0.924 ± 0.3192.151 ± 0.034
UGradSL+EL2N1.042 ± 0.2092.362 ± 0.061
UGradSL+GraNd0.950 ± 0.2673.405 ± 0.069
UGradSL+InfoBatch1.548 ± 0.4862.252 ± 0.009
UGradSL+Random2.052 ± 0.2181.575 ± 0.001
UGradSL+Fixed initial scores0.941 ± 0.1941.630 ± 0.001
RLFULL1.326 ± 0.1701.666 ± 0.000
RLDSS1.808 ± 0.3201.461 ± 0.057
RLDSS+1.300 ± 0.1501.474 ± 0.034
RLEL2N1.828 ± 0.2291.423 ± 0.068
RLGraNd2.133 ± 0.3172.366 ± 0.049
RLInfoBatch1.923 ± 1.2941.250 ± 0.010
RLRandom1.095 ± 0.1580.875 ± 0.000
RLFixed initial scores1.046 ± 0.1350.930 ± 0.000
SalUnFULL1.669 ± 0.4921.683 ± 0.000
SalUnDSS1.623 ± 0.1751.457 ± 0.025
SalUnDSS+1.376 ± 0.1501.510 ± 0.022
SalUnEL2N1.689 ± 0.2321.449 ± 0.049
SalUnGraNd1.941 ± 0.1942.411 ± 0.074
SalUnInfoBatch4.273 ± 7.6991.281 ± 0.020
SalUnRandom1.243 ± 0.1820.891 ± 0.000
SalUnFixed initial scores1.041 ± 0.1890.947 ± 0.000
AMUNFULL0.950 ± 0.1552.307 ± 0.007
AMUNDSS0.850 ± 0.2791.346 ± 0.027
AMUNDSS+1.020 ± 0.5491.441 ± 0.044
AMUNEL2N1.046 ± 0.3231.374 ± 0.049
AMUNGraNd1.069 ± 0.5412.461 ± 0.021
AMUNInfoBatch0.883 ± 0.2231.713 ± 0.007
AMUNRandom1.071 ± 0.2281.436 ± 0.007
AMUNFixed initial scores1.004 ± 0.2141.492 ± 0.007

Table R2. CIFAR-100 / VGG16, random forgetting. Mean ± sample SD across ten paired requests at the reported endpoint. Avg. Gap is in percentage points. The complete trajectories appear in Figure 12 of the revised manuscript (Appendix B.2).

MethodSelectionAvg. GapPFLOPs
NegGrad+FULL14.471 ± 0.4271.678 ± 0.000
NegGrad+DSS12.109 ± 2.9191.143 ± 0.102
NegGrad+DSS+12.223 ± 1.6201.201 ± 0.067
NegGrad+EL2N14.098 ± 1.3521.093 ± 0.041
NegGrad+GraNd14.244 ± 0.6941.780 ± 0.021
NegGrad+InfoBatch11.569 ± 1.6311.120 ± 0.003
NegGrad+Random13.804 ± 1.7431.255 ± 0.003
NegGrad+Fixed initial scores13.622 ± 1.0851.284 ± 0.005
RLFULL6.666 ± 0.2470.941 ± 0.000
RLDSS6.323 ± 1.6000.956 ± 0.021
RLDSS+12.916 ± 17.2230.900 ± 0.013
RLEL2N4.884 ± 0.6640.930 ± 0.009
RLGraNd11.537 ± 21.4111.474 ± 0.113
RLInfoBatch9.404 ± 0.4110.762 ± 0.010
RLRandom12.427 ± 14.9880.518 ± 0.000
RLFixed initial scores7.205 ± 0.3850.549 ± 0.000

The Introduction clarifies the distinction between the base objective and the examples used to compute updates. Section 5.2 (Effect of the selection rule) and Appendix A.3 explain the score replacements. The expanded comparisons have been added to Appendix B.2 as Figures 11 and 12, with descriptions of the results alongside the figures.

W2. Thank you for pointing this out. We restricted the analysis to binary logistic regression with Gaussian inputs so that the source-update comparison and its proof could be completed under explicit assumptions. Its purpose is to explain when an update from a forget example moves predictions closer to the retrained model. Applicability across different base unlearning methods is evaluated empirically. The theoretical result is conditional on the stated model, rather than a general guarantee for every algorithm in the framework.

Section 4.1 states the logistic model, Gaussian inputs, source conditions and conditional comparison in Proposition 1. Appendix C gives the corresponding assumptions, proofs and extensions. The Introduction now explicitly describes the analysis as binary logistic regression with Gaussian inputs.

W3. Tables 1–2 fix the learning rate and ten-epoch training period across FULL, DSS and DSS+ within each base method. This controls the settings other than selection. Because the selected policies use fewer examples, they can receive fewer updates and reach a different stage of training at that endpoint.

Figure 3 includes training beyond ten epochs to examine this issue. In some settings, additional training brings selected variants closer to the corresponding FULL reference while their cumulative computation remains lower. The endpoint tables and extended curves therefore address both the result at a common training period and the computation required as training continues.

Our experiments concern base unlearning methods that use both retain and forget data. Methods that use only retain data are outside this empirical scope. The retain-only selection comparison in our additional ablation keeps the forget objective and tests which data pool needs selection, a different question from evaluating a retain-only unlearning method.

The common training period is specified in Section 5.1. Section 5.2 (Performance) discusses additional training using Figure 3. The separate selection-side ablation has been added to Section 5.2 (Which set to select), Table 4. This ablation retains both loss terms and does not constitute an evaluation of retain-only unlearning algorithms.

W4. We agree that these results do not establish general computational efficiency for LLMs. The experiment was an extension of the vision selection procedure to answer tokens. We consider LLM token selection better treated as a separate study and have removed the LLM experiments and associated claims from the revised manuscript.

Our LLM experiments and associated claims have been removed from the Abstract, Introduction, experimental setup and results, Conclusion, and token-selection appendix. References to other work on LLM unlearning remain in Related Work.

W5. We use an initial selection fraction of 5% and refresh after every epoch as the default configuration. For the main tables, we fix the learning rate and end all policies at ten epochs. In some settings, we increase the initial fraction so that training can make sufficient progress within this common period.

A smaller fraction reduces computation per epoch but can require more epochs. Figure 5 shows the effect of the initial fraction, and Figure 3 shows the effect of extended training. The reported fraction adjustments should therefore be read together with the common training period used in the tables. The refresh-interval experiment separately examines the effect of retaining the same selected subset for more epochs before rescoring.

Section 5.3 (Discussion) now explains the starting configuration, why larger fractions are used in some settings, and how training duration and refresh frequency should be considered together. Section 5.2 reports fraction and interval sensitivity in Figures 5 and 6. Appendix A.2 records the fractions used in the experiments.

Reviewer_yPhTOpenReview ↗

Official Comment by Reviewer yPhT

I thank the authors for their responses. However, I have concerns about the novelty of this work. Furthermore, the performance of the proposed algorithm compared with other methods is not convincing, both in terms of the average gap and PFLOPs. Therefore, I maintain my original score.

Reviewer_qs2rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper is investigating the machine unlearning setup with removal of random class labels. The objective is to find better ways to identify retain and forget points in order to speed up the unlearning processes further, without compromising the performance. The method is called Dynamic Source Selection (DSS) or DSS+ depending on the exact selection scheme. The selection method can be applied to existing machine unlearning algorithm as a wrapper. Experiments are performed on CIFAR-10, CIFAR-100 and also on TOFU experiments. Overall the results are convincing and the method demonstrate speedups for most methods and configurations.

Critical strengths

The proposed method is well explained, the figure 1 gives a good overview and algorithm 1 outlines the process nicely. The experiments are convincing, and the method clearly has merit. For this reason, I think the paper good could be a good addition to the conference, assuming the authors can fix the weaknesses.

Critical weaknesses

he primary weakness of the paper is the clarity. The paper is not explaining the reference material properly which makes the paper difficult to read, especially for people new or outside machine unlearning. I grant that most of the terms are known and/or can be identified by searching the web, however a paper should be self-contained. Here is a list of things that I think should be explained / defined cited:

  • L20: DSS+ - implicit naming - it's not clear when you read the abstract if this is a typo or another name
  • L45: Gaussian logistic analysis, what is this, add reference (do you mean logistic regression?)
  • L58: Ref to CIFAR
  • L61: Add ref to TOFU
  • L91: Since you use EL2N, it would probably be a good idea to describe it.
  • L93-95: This sentence is not related work, this is your work, so should be moved elsewhere.
  • L101: GraNd+EL2N not defined.
  • L103: WMDP and MUSE are not defined.
  • L232: of-fline -> off-line
  • L300: here TOFU is cited, but it should happen in L61
  • The figures (figure 3-7) are actually a bit hard to read - probably adjusting the line width, zooming or color scheme will make them better. E.g on Fig 4 the colors from 1%->90% could follow a palette.
  • Table 1+2 seems to be using a pretty small font size making it hard to read on a printout.
Questions for authors
  1. For equation 2, it is not clear to me what rule v and u play. Can you elaborate on this and give some intution?
  2. Contemplating on table 1+2, the goal of DSS and DSS+ is to speed up unlearning, without changing the outcome of the original methods, correct? If so, I'm wondering how the results should be reported so they are as easy to understand as possible. Should I compare DSS / DSS+ performance with the original method, or the retrained model ?
2 comments
AuthorsOpenReview ↗

Response to Reviewer qs2r

Thank you for your helpful comments. We have revised the manuscript accordingly and address your questions below.

Q1. In Equation 2, θR=βv\theta_R=\beta v is the parameter vector of the model retrained on the retain set, and θ0−θR=γu\theta_0-\theta_R=\gamma u is its difference from the initial model. The unit vectors vv and uu therefore represent the retrained model's direction and the direction of the initial discrepancy, respectively. The scalars β\beta and γ\gamma specify their magnitudes.

The assumption u⊤v=0u^\top v=0 separates these two directions. An update component along −u-u corrects the initial discrepancy, while a component along vv changes the retrained model's signal. This decomposition lets the analysis distinguish these effects when comparing updates from two forget examples.

Section 4.1, immediately below Equation (2), defines the directions vv and uu and the magnitudes β\beta and γ\gamma. The same subsection states their orthogonality and the equal-input-norm condition used in the update comparison.

Q2. The model retrained on the retain set is the quality reference. Avg. Gap measures the difference from that reference. FULL provides the performance and computational-cost baseline for the underlying unlearning method, allowing us to assess what changes when DSS or DSS+ is applied.

The relative benefit of DSS and DSS+ depends on the method and setting. For example, under CIFAR-10 random forgetting, RL with DSS+ has an Avg. Gap of 1.31 pp, compared with 1.82 pp for DSS (Table 1, left). In the same setting, NegGrad+ with DSS has a lower Avg. Gap and runtime than DSS+. The selection-transition analysis also shows that many RL examples remain selected across consecutive epochs, a pattern that DSS+ accommodates without rescoring previously selected examples.

Tables 1–2 use the same learning rate and ten epochs across policies within each method. Because DSS and DSS+ use fewer examples, equal epochs do not imply equal numbers of updates. The extended trajectories in Figure 3 show how the discrepancy from the retrained model, measured by Avg. Gap, changes with additional training and cumulative computation. The quality reference remains the retrained model, and FULL provides a baseline for assessing the effect of selection on quality and cost.

Section 5.1 (Baseline Methods and Evaluation Metrics) specifies the comparison baseline, retrained reference and common training period. Section 5.2 discusses the extended trajectories in Figure 3 and the differences between DSS and DSS+. The newly added runtime breakdown appears in Section 5.2, Table 3.

I appreciate the answers. Having reflected on the other reviewers criticism of novelty, I have decided to keep the score of 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.