acceptodds
Under review as a conference paper at ICLR 2027

Geometry of Forgetting: Representation Flux in Continual Learning

Abstract

Catastrophic forgetting remains a fundamental obstacle to continual learn- ing, where neural networks lose previously acquired knowledge while learn- ing new tasks. Existing methods primarily mitigate forgetting through pa- rameter regularization or experience replay, yet the representation-space dy- namics associated with forgetting remain less well understood. In this work, we investigate the evolution of latent representations during sequential learning and introduce representation flux, a geometric quantity that mea- sures sample-level representation displacement across training. We show that representation flux is strongly associated with catastrophic forgetting across multiple continual learning benchmarks, while temporal analyses provide evidence that elevated flux can precede subsequent performance degradation. Larger representation displacement is also associated with greater confidence degradation, and complementary geometric properties of representation transitions provide additional information about sample- level forgetting. Motivated by these observations, we propose FlowLess-R, a simple representation-space regularization method that constrains replay- sample representations relative to stored reference representations while al- lowing continued learning of new tasks. FlowLess-R is architecture-agnostic and can be integrated into existing replay-based continual learning methods by adding a representation-matching term to the training objective. Ex- periments on SplitMNIST, SplitFashionMNIST, SplitCIFAR10, and Split- TinyImageNet demonstrate improvements in final average accuracy and reductions in catastrophic forgetting when FlowLess-R is combined with ER, DER++, and ER-ACE. Together, our results identify representation flux as an informative geometric marker of forgetting and show that sta- bilizing latent representations provides a simple and effective strategy for mitigating catastrophic forgetting.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Continual Learning, Catastrophic Forgetting, Representation Learning, Representation Geometry, Representation Drift, Experience Replay, Stability–Plasticity

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hXB2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies sample-level displacement of latent representations between epochs ("representation flux") in replay-based continual learning. It reports that flux correlates with, and tends to precede, forgetting events. It then proposes FlowLess-R, an L2 penalty between the current features of buffer samples and features stored when those samples entered memory. FlowLess-R is added to ER, DER++ and ER-ACE and evalutated Split MNIST/FashionMNIST/CIFAR-10/TinyImageNet.

Critical strengths

The addressed problem is relevant and the method is simple.

Critical weaknesses
  1. That forgetting stems from drift of old representations is the premise of all feature-distillation methods in CL, and it has been measured before (Davari et al., CVPR 2022; Caccia et al., ICLR 2022). The flux analysis restates this premise at sample level and remains correlational.

  2. L2 feature matching against past features is standard: Jung et al. (AAAI 2018), Michieli & Zanuttigh (ICCVW 2019), LUCIR (CVPR 2019), PODNet (ECCV 2020), BFP (CVPR 2023), EFC (ICLR 2024). Using targets stored at buffer insertion is what DER (NeurIPS 2020) already does with logits. Thus, the novelty of the method appears to be very limited.

  3. Replay samples are selected uniformly at random, with no comparison to or combination with coreset selection methods such as GSS, Rainbow Memory or OCS. This is a natural experiment given the sample-level analysis.

Questions for authors

N/A

3 comments
AuthorsOpenReview ↗

Response to Reviewer hXB2

We thank the reviewer for the constructive feedback and relevant references. We address the concerns below.

1. Representation flux and prior drift analyses

We acknowledge that representation drift has been studied previously (Davari et al., 2022; Caccia et al., 2022). Our contribution is not the discovery of drift, but its sample-level temporal characterization and relationship to forgetting. We examine whether elevated representation flux precedes forgetting events and whether it provides information beyond historical displacement.

The revised manuscript compares representation flux against historical L2, cosine, and normalized drift (Section 4, Table 2, p. 6). Adding flux to combined historical drift improves concurrent forgetting-detection ROC-AUC by 0.033 on SplitMNIST, 0.049 on SplitFashionMNIST, 0.230 on SplitCIFAR10, and 0.100 on SplitTinyImageNet. Paired 95% confidence intervals for these improvements exclude zero across all four benchmarks (Appendix B.1, Table 6, p. 19). These results demonstrate that flux provides complementary discriminative information beyond historical drift.

The distinction between instantaneous representation flux and accumulated historical drift is further formalized in Appendix A.6 (Proposition 3, pp. 17–18), showing that historical drift does not uniquely determine instantaneous flux.

2. FlowLess-R and feature distillation

We agree that L2 feature matching is well established in continual learning (Jung et al., 2018; Michieli & Zanuttigh, 2019; LUCIR, 2019; PODNet, 2020; BFP, 2023; EFC, 2024), as is storing historical targets, exemplified by DER (2020). We do not claim novelty for either technique.

Our primary contribution is the geometric and temporal analysis of catastrophic forgetting. FlowLess-R serves as a simple, controlled intervention illustrating how this analysis can motivate representation stabilization, rather than as a fundamentally new regularization method.

To strengthen the empirical evaluation, we added comparisons with frozen-teacher distillation, FER-style feature-and-logit preservation, and an adapted PODNet implementation under shared class-incremental protocols. Replay tensor storage is matched for the first three benchmarks; frozen-teacher parameters are excluded from this budget.

Final average accuracy (%, mean ± SD):

MethodSplitMNISTSplitFashionMNISTSplitCIFAR10SplitTinyImageNet
ER72.98 ± 1.4564.73 ± 0.9238.10 ± 1.6616.33 ± 0.94
Frozen teacher75.29 ± 1.4166.38 ± 1.9637.45 ± 3.6317.70 ± 2.27
FER-style76.96 ± 1.5272.22 ± 0.8636.17 ± 1.9216.78 ± 1.03
PODNet-full (adapted)69.95 ± 3.0164.19 ± 2.2842.97 ± 3.4116.11 ± 0.53
FlowLess-R (epoch-refresh)72.79 ± 1.6867.15 ± 1.6439.66 ± 0.5719.18 ± 0.80
FlowLess-R (fixed anchors)79.00 ± 1.5771.27 ± 1.5839.50 ± 0.9318.21 ± 1.59

Source: Appendix D.5, Table 10, p. 29, with implementation details on pp. 28–29.

Fixed-anchor FlowLess-R improves over ER by 6.02, 6.54, 1.40, and 1.88 percentage points, respectively. It achieves the highest accuracy on SplitMNIST, while FER-style leads on SplitFashionMNIST, adapted PODNet on SplitCIFAR10, and epoch-refresh FlowLess-R on SplitTinyImageNet. Thus, no method universally dominates.

These comparisons support the practical value of FlowLess-R while reinforcing our central claim: the paper's primary novelty lies in characterizing the geometry and temporal dynamics of forgetting, with representation regularization serving as an illustrative intervention motivated by the analysis.

3. Replay selection strategies

Replay selection and representation stabilization address orthogonal aspects of continual learning: GSS, Rainbow Memory, and OCS determine which samples are retained, whereas FlowLess-R constrains how their representations evolve during subsequent training. We use uniform random replay to isolate the effect of representation stabilization independently of selection policies. FlowLess-R is compatible in principle with alternative replay-selection strategies; evaluating their combinations is beyond the scope of this study.

4. Soundness assessment

We respectfully request clarification of the soundness rating of 1. The stated concerns primarily address novelty and experimental scope. Could the reviewer identify any specific mathematical, methodological, or experimental validity issue that motivates this rating? We would welcome the opportunity to address it directly.

We thank the reviewer again for their consideration.

Reviewer_hXB2OpenReview ↗

Official Comment By Reviewer hXB2

Thanks for the detailed response and the additional experiments.

My main concern remains. The rebuttal confirms that the regularizer is not novel, so the contribution rests on the flux analysis. However, the method does not use flux: it penalizes accumulated drift from anchors stored at insertion, which your new Table 2 shows to be less informative. One question: why not regularize flux directly?

On soundness, my rating reflects this gap between analysis and method and the evaluation issue that changed the main results during the rebuttal. I think the paper needs a major revision and a new round of review, so I keep my score.

AuthorsOpenReview ↗

Reponse to Official Comment By Reviewer hXB2

We thank the reviewer for the prompt response and clarification.

We agree that FlowLess-R constrains accumulated displacement rather than instantaneous flux. This distinction is intentional: direct flux regularization discourages consecutive representation changes but permits gradual accumulated drift, whereas fixed anchors preserve historical representations while allowing local adaptation, thereby providing a more persistent constraint against cumulative representation drift.

The two displacement measures are mathematically related through the triangle inequality, although controlling accumulated drift does not necessarily control instantaneous flux.

Importantly, we also investigate task-wise and epoch-wise anchor refreshing in the revised manuscript (Section 5, Table 5), which more closely approximates regularization of shorter-term representation changes.

Table 5: Anchor-refresh ablation (final average accuracy, %; ER + FlowLess-R).

Anchor policySplitMNISTSplitFashionMNISTSplitCIFAR10SplitTinyImageNet
Fixed79.5670.7940.1618.21
Task-wise79.5268.6939.6517.18
Epoch-wise73.8866.2640.2319.18

These results suggest that neither fixed nor frequently refreshed anchors universally dominate. Nevertheless, epoch-wise refreshing is not equivalent to directly minimizing instantaneous flux.

Our primary contribution remains the sample-level geometric and temporal characterization of forgetting. In particular, Table 2 of the revised manuscript demonstrates that flux provides complementary information beyond historical drift.

Table 2: Representation flux versus historical drift (concurrent forgetting-detection ROC-AUC).

FeaturesSplitMNISTSplitFashionMNISTSplitCIFAR10SplitTinyImageNet
Historical L2 Drift0.6430.6170.5250.576
Historical Cosine Drift0.6320.5520.5190.594
Normalized Historical L20.6530.6540.5260.597
Historical Drift (Combined)0.7420.7150.5360.621
Representation Flux0.7310.7020.7660.673
Historical Drift + Flux0.7760.7640.7660.721

Adding flux to combined historical drift improves ROC-AUC by 0.033, 0.049, 0.230, and 0.100, respectively, across the four benchmarks. These findings support the distinction between instantaneous flux and accumulated drift, demonstrating that flux captures complementary information about forgetting.

FlowLess-R serves as an illustrative stabilization intervention motivated by this geometric analysis, rather than a direct optimization of flux.

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

The paper studies catastrophic forgetting in continual learning. It introduces representation flux, which measures how much the latent representation of individual samples changes between training steps. The authors show empirically that larger representation change is associated with greater forgetting and confidence degradation, and provide temporal evidence that increased flux can precede performance degradation. Based on these observations, the authors propose FlowLess-R, a regularization method that stores historical representations of replay samples and penalizes their deviation from newer representations during later training. Experiments show that FlowLess-R reduces forgetting and often improves final accuracy.

Critical strengths

The paper asks an interesting question beyond benchmark performance, and the analysis is interesting. Another positive point that is quite valuable is the use of multiple seeds and reporting statistical testing with Holm–Bonferroni correction across tested \lambda values. Plus the ablations investigate a wide range of hyperparameters.

Critical weaknesses

The main criticism of the paper is its novelty. There are a lot of other recent methods that use similar techniques to improve the forgetting or try to explain it with newer methods/analysis. However, since the experiments are all on simpler datasets, smaller models (Resnet18) and older continual models, the novelty becomes less clear. Also, we can observe that on the more challenging dataset TinyImageNet the gain is not that much. Given the more recent trend of using large foundation models, could the authors relate their findings to those models and show that for larger models their findings are still valuable?

Questions for authors

How much additional GPU/CPU memory does FlowLess-R use for a given improvement? Does FlowLess-R still outperform standard replay under an equal total memory budget when the storage required for historical feature vectors is counted? Why does flux alone become close to uninformative on SplitTinyImageNet while the combined geometric descriptors retain some predictive ability? Do the authors expect the usefulness of flux to decrease further with task/model complexity? How does FlowLess-R compare experimentally with existing representation-preservation/distillation approaches, particularly methods using stored historical prototypes? Does the age of the stored representation matter? Is anchoring to the representation at buffer insertion preferable to periodically updating the reference? How is representation flux conceptually and empirically different from existing measures of representational shift/discrepancy in continual learning?

2 comments
AuthorsOpenReview ↗

Review to Enjr (first part)

Response to Reviewer Enjr

We thank the reviewer for the constructive feedback and for recognizing the statistical rigor and motivation of our analysis. We address the concerns below in two blocks.

1. Memory overhead and equal-budget comparison

We evaluated FlowLess-R under equal total-memory budgets, reducing replay capacity to account for historical feature storage (Tables 3 and 7). Performance remains comparable to standard replay, indicating that FlowLess-R remains competitive when its additional storage requirements are included. Historical feature storage introduces O(Md) overhead for M samples and feature dimension d.

2. Representation flux on SplitTinyImageNet and increasing complexity

We thank the reviewer for highlighting the unexpectedly low ROC-AUC. Upon investigation, we identified an evaluation bug: unintended data augmentation distorted the representation statistics. After correction, flux alone achieves ROC-AUC 0.673, while combined geometric descriptors achieve 0.819 (Section 3.3, Table 1, p. 6).

We additionally compare flux against historical representation-drift measures for concurrent forgetting detection (Table 2, p. 6).

ROC-AUC (mean ± SD):

FeaturesSplitMNISTSplitFashionMNISTSplitCIFAR10SplitTinyImageNet
Historical L2 Drift0.643 ± 0.0230.617 ± 0.0250.525 ± 0.0050.576 ± 0.008
Historical Cosine Drift0.632 ± 0.0420.552 ± 0.0520.519 ± 0.0040.594 ± 0.010
Normalized Historical L20.653 ± 0.0400.654 ± 0.0300.526 ± 0.0050.597 ± 0.003
Historical Drift (Combined)0.742 ± 0.0120.715 ± 0.0090.536 ± 0.0050.621 ± 0.008
Representation Flux0.731 ± 0.0040.702 ± 0.0170.766 ± 0.0020.673 ± 0.002
Historical Drift + Flux0.776 ± 0.0110.764 ± 0.0140.766 ± 0.0020.721 ± 0.001

Results use leave-one-seed-out evaluation over five runs for SplitMNIST, SplitFashionMNIST, and SplitCIFAR10, and three runs for SplitTinyImageNet.

Adding flux to combined historical drift improves ROC-AUC by approximately 0.034, 0.049, 0.230, and 0.100 across the four benchmarks, respectively. Flux therefore provides complementary information beyond historical displacement.

The corrected TinyImageNet results do not support the original near-chance interpretation. Although flux informativeness varies across benchmarks, our experiments do not establish a monotonic decline with increasing complexity.

3. Comparison with representation-preservation methods

We added comparisons against frozen-teacher distillation, FER-style preservation, and adapted PODNet (Appendix D.5, Table 10, p. 29).

Final average accuracy (%, mean ± SD):

MethodSplitMNISTSplitFashionMNISTSplitCIFAR10SplitTinyImageNet
ER72.98 ± 1.4564.73 ± 0.9238.10 ± 1.6616.33 ± 0.94
Frozen teacher75.29 ± 1.4166.38 ± 1.9637.45 ± 3.6317.70 ± 2.27
FER-style76.96 ± 1.5272.22 ± 0.8636.17 ± 1.9216.78 ± 1.03
PODNet-full (adapted)69.95 ± 3.0164.19 ± 2.2842.97 ± 3.4116.11 ± 0.53
FlowLess-R (fixed)79.00 ± 1.5771.27 ± 1.5839.50 ± 0.9318.21 ± 1.59
FlowLess-R (epoch)72.79 ± 1.6867.15 ± 1.6439.66 ± 0.5719.18 ± 0.80

No method consistently dominates: FlowLess-R leads on SplitMNIST and SplitTinyImageNet, FER-style on SplitFashionMNIST, and adapted PODNet-full on SplitCIFAR10.

Our primary novelty lies in the observation and temporal characterization of sample-level representation flux and its relationship to forgetting, not in the regularization mechanism itself. Historical feature matching is established; FlowLess-R serves as a simple intervention motivated by our geometric analysis.

AuthorsOpenReview ↗

Review to Enjr (part 2)

4. Historical-anchor age and refresh frequency

We conducted additional experiments comparing fixed, task-wise, and epoch-wise anchor refresh (Section 5.5, Table 5, p. 9).

Final average accuracy (%, mean ± SD):

DatasetFixedTask-wiseEpoch-wise
SplitMNIST79.56 ± 1.6379.52 ± 1.4773.88 ± 1.13
SplitFashionMNIST70.79 ± 1.0668.69 ± 1.5266.26 ± 1.12
SplitCIFAR1040.16 ± 2.0539.65 ± 2.3140.23 ± 1.71
SplitTinyImageNet18.21 ± 1.5917.18 ± 1.9419.18 ± 0.80

Fixed anchors perform best on SplitMNIST and SplitFashionMNIST, while epoch-wise refresh performs best on SplitTinyImageNet. Differences on SplitCIFAR10 are small.

We also investigated whether longer training favors more frequent anchor updates (Appendix D.4, Table 9, p. 28). Across 5, 10, 20, and 40 epochs per task on SplitFashionMNIST, fixed anchors consistently achieved the highest accuracy and lowest forgetting. Thus, we found no evidence supporting this hypothesis in the tested setting. The preferred refresh policy appears dataset-dependent.

5. Novelty relative to existing representation-shift measures

We acknowledge that representation drift and L2 feature preservation are established concepts. Our primary contribution is the sample-level temporal characterization of forgetting geometry, rather than a fundamentally new regularization technique. The revised manuscript compares representation flux with historical L2, cosine, and normalized drift (Section 4, Table 2, p. 6). Adding flux to combined historical drift improves concurrent forgetting-detection ROC-AUC by approximately 0.034, 0.049, 0.230, and 0.100 across SplitMNIST, SplitFashionMNIST, SplitCIFAR10, and SplitTinyImageNet, respectively. Paired 95% confidence intervals exclude zero (Appendix B.1, Table 6, p. 19).

Appendix A.6 (Proposition 3, pp. 17–18) further establishes that historical displacement does not uniquely determine instantaneous flux.

Our temporal analyses additionally investigate whether elevated flux precedes forgetting. We interpret this relationship as predictive rather than causal. FlowLess-R serves as a simple intervention motivated by this geometric analysis.

6. Larger architectures and foundation models

Evaluation on ResNet-50 and ViTs would strengthen the study, but computational and rebuttal-time constraints prevent rigorous multi-seed experiments.

We thank the reviewer for the constructive feedback. The revised manuscript includes corrected results, additional baseline comparisons, anchor-refresh ablations, and clarifications of the paper's contributions and limitations. All references refer to the revised version.

Reviewer_ps47OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the problem of catastrophic forgetting in continual learning. The authors firstly examine whether the change in representation geometry could be a valid metric to measure catastrophic forgetting. They then propose a measure named "Representation Flux", which measures the change of the representation of a sample xx between epochs, and they provide the empirical evidence to support their claims. Inspired by this, the authors propose a novel method, which uses a replay buffer to calculate the "Representation Flux" and then uses it as the penalty. They conduct experiments on different datasets with various architectures to empirically verify the effectiveness of the proposed method.

Critical strengths
  • The proposed method is simple but effective. They require a replay buffer to estimate the "Representation Flux" at each epoch, and it is flexible enough to be combined with other methods.
  • This proposed method is well-motivated to me. The authors intuitively connect forgetting with representation drift, and then find a proper metric to reflect the representation drift. They provide several experiments to support their finding, which helps me to understand the motivation behind the method.
Critical weaknesses
  • The latent representation of the sample z~i\tilde{z}_i seems not to be updated during the training process after the sample is added into the replay buffer, which seems to be inconsistent with the definition of "Representation Flux". Moreover, In Table 1, the performance of Flux alone on SplitTinyImageNet is significantly lower than that on other datasets, with the AUC close to 0.5, which implies that Flux may be invalid for measuring the forgetting per sample in this complex dataset.
  • The critical component of the proposed method is the construction of a valid and informative replay buffer. A discussion of how to select samples and how to identify the proper size of the replay buffer should be included in this paper. Moreover, the cost of using the replay buffer seems to be grows as the number of tasks.
Questions for authors
  • Why the latent representation of the sample z~i\tilde{z}_i is not updated in each epoch. How do the authors ensure that the stored representation does not become stale?
  • The experiments seem to be insufficient. The biggest architecture used is ResNet-18. I am concerned about the performance of the proposed method on bigger NNs, e.g., ResNet-50, ViT-B, ViT-L.
1 comment
AuthorsOpenReview ↗

Response to Reviewer ps47

We thank the reviewer for the positive assessment and constructive suggestions. We address each concern below.

1. Historical anchors versus representation flux

Representation flux measures displacement between consecutive epochs, whereas FlowLess-R penalizes deviation from historical representations stored when samples enter the replay buffer (Section 4.1). Thus, FlowLess-R does not directly penalize consecutive-epoch flux. Current representations continue to evolve, while historical anchors are intentionally preserved to constrain accumulated drift.

To investigate anchor staleness, we conducted additional experiments comparing fixed, task-wise, and epoch-wise refresh policies (Section 5.5, Table 5, p. 9).

DatasetFixedTask-wiseEpoch-wise
SplitMNIST79.56 ± 1.6379.52 ± 1.4773.88 ± 1.13
SplitFashionMNIST70.79 ± 1.0668.69 ± 1.5266.26 ± 1.12
SplitCIFAR1040.16 ± 2.0539.65 ± 2.3140.23 ± 1.71
SplitTinyImageNet18.21 ± 1.5917.18 ± 1.9419.18 ± 0.80

Final average accuracy (%, mean ± SD).

Fixed anchors perform best on SplitMNIST and SplitFashionMNIST, while epoch-wise refresh performs best on SplitTinyImageNet. Differences on SplitCIFAR10 are small. We additionally tested whether longer training favors more frequent anchor updates (Appendix D.4, Table 9, p. 28). Across 5, 10, 20, and 40 epochs per task on SplitFashionMNIST, fixed anchors consistently achieved the highest accuracy and lowest forgetting. Thus, the hypothesis that longer training necessarily reduces the effectiveness of fixed anchors was not supported in this setting. These findings indicate a dataset-dependent stability–plasticity trade-off rather than a universally optimal refresh policy.

2. Representation flux on SplitTinyImageNet

We thank the reviewer for highlighting the unexpectedly low ROC-AUC. Upon investigation, we identified a bug in our evaluation pipeline: unintended data augmentation distorted the representation statistics on SplitTinyImageNet. After correcting this issue, representation flux alone achieves ROC-AUC 0.673, substantially above chance (0.5). Combining geometric descriptors increases ROC-AUC to 0.819 on the same dataset (Section 3.3, Table 1, p. 6). These corrected results demonstrate that representation flux provides meaningful information about sample-level forgetting on SplitTinyImageNet, while complementary geometric descriptors further improve discrimination. We appreciate the reviewer's observation, which helped us identify and resolve the issue.

3. Replay-buffer selection, size, and computational cost

FlowLess-R does not require a specialized replay-selection mechanism. We deliberately use uniform random replay to isolate representation stabilization from sample-selection strategies.

We investigate different replay capacities in Section 5.3 and Appendix D.2 and provide memory-matched comparisons in Tables 3 and 7. While these experiments characterize the influence of replay capacity, identifying an optimal buffer size or selection policy remains dataset-dependent.

We also added comparisons with FER-style feature preservation and adapted PODNet-full (Appendix D.5, Table 10, p. 29). No method consistently dominates: FlowLess-R achieves the highest accuracy on SplitMNIST and SplitTinyImageNet, FER-style on SplitFashionMNIST, and adapted PODNet-full on SplitCIFAR10. Evaluating combinations with more sophisticated replay-selection policies remains future work.

FlowLess-R requires O(Md) additional storage for M replay samples with representation dimension d and O(B_r d) regularization computation per replay batch (Section 4.2). Under a fixed total replay-memory budget, overhead does not inherently grow with the number of tasks.

4. Generalization to larger architectures

We agree that evaluation on ResNet-50 and especially Vision Transformers (ViT-B and ViT-L) would strengthen the study, as ViTs represent a fundamentally different architectural family from CNNs. However, limited computational resources and the short rebuttal period prevent rigorous multi-seed experiments.

We thank the reviewer for the constructive feedback, particularly for helping identify the SplitTinyImageNet evaluation issue. We have updated the manuscript accordingly, incorporating the corrected results, additional experiments, and methodological clarifications. All section, table, and page references in this response refer to the revised manuscript.

Comments on the paper

1 comment

Dear Reviewers,

Please read the rebuttals and comment whether they address your concerns. When appropriate please adjust your scores.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.