acceptodds
Under review as a conference paper at ICLR 2027

Understanding Robustness to Noisy Supervision in Graph Neural Networks

Abstract

Graph Neural Networks (GNNs) are vulnerable to noisy supervision, yet the role of architecture in mitigating this effect remains poorly understood. We study a modified version of Jumping Knowledge (JK) that preserves direct access to pre-propagation node representations by combining the initial and final representations at readout, while discarding intermediate propagation states. We theoretically show that this construction preserves information that message passing may make inaccessible to the prediction head, and characterize its advantage in terms of the prediction deficit of the propagated representation and the recoverability of this deficit from the pre-propagation representation. Across four GNN architectures, four datasets, three noise mechanisms, and three corruption levels, we find that JK improves accuracy under label noise by 4.3 points (around 7.0% relative) on average, with its benefit increasing with corruption in the regimes where information made inaccessible through propagation remains recoverable from the initial representation. Our analysis shows that label noise enlarges the prediction deficit across the corruption range considered, while recoverability remains comparatively stable, identifying recoverable prediction deficit as a key factor governing when representation preservation improves robustness to noisy supervision.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Noisy Labels, Robustness, Jumping Knowledge, Graph Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VKhSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper analyzes whether preserving pre-propagation node representations (w/ simplified JK readout) improves GNN robustness to noisy labels. The author theoretically characterizes the benefit using prediction deficit and recoverability.

Critical strengths

The proposed method is straightforward and easy to follow

The theorem shows the value of preserving information

Experimental analysis is broad

Critical weaknesses

Concatenating initial and final representations closely resembles existing JK and skip-connection designs. In addition, the theoretical results follow from feature concatenation and standard linear projection identities

The analysis assumes fixed representations. Thus, it is ambiguous how the proposed method improves test accuracy under noisy supervision

Empirical gain is marginal on Amazon-Ratings. GraphSAGE almost converges to 0. These results make it difficult to isolate noise-specific robustness

Questions for authors

Please see above

1 comment
AuthorsOpenReview ↗

Response to Reviewer VKhS

  • Novelty. Our contribution is not a new JK architecture, but a theoretical and empirical characterization of representation preservation as a robustness mechanism under noisy supervision. First, JK differs from layer-wise residual connections: residuals combine representations during propagation, whereas JK guarantees direct access to H(0)H^{(0)} at readout. Importantly, GCN*, GAT* and SAGE* already employ learned residual connections, yet JK provides substantial improvements for GCN* and GAT*. The limited gains observed for SAGE* further highlight that the benefit of JK depends on how each backbone preserves node-specific information. Second, while our theoretical results build on standard linear projection identities, their contribution is to characterize when representation preservation improves prediction. The exact decomposition ΔJK∗=D(ρ) Γ(ρ)\Delta^*_{\mathrm{JK}} = D(\rho)\,\Gamma(\rho) distinguishes prediction deficit from recoverability, showing that losing predictive information alone is insufficient for JK to help. In particular, even a large prediction deficit yields no JK advantage when the missing signal is not recoverable from H(0)H^{(0)}, a distinction that cannot be established from predictor-class containment alone. Recoverability is approximately 0.97 on BlogCatalog, 0.55 on Flickr, 0.31 on Roman Empire, and 0.10 on Amazon-Ratings, matching the ordering of probe gains at every nonzero corruption rate. Notably, Amazon-Ratings exhibits the largest prediction deficit but negligible probe gains, demonstrating why information loss alone cannot explain JK's effectiveness. Unlike prior JK and skip-connection studies focused on heterophily, depth, or oversmoothing, our work connects representation preservation to noisy supervision and identifies the conditions governing its effectiveness.

  • Fixed representations. Representations are fixed only during the analysis, not during training. Both H(0)H^{(0)} and H(T)H^{(T)} are learned under noisy supervision at rate ρ\rho, so noise affects D(ρ)D(\rho) and Γ(ρ)\Gamma(\rho). Our theory characterizes the clean predictive signal accessible to a linear readout, while linear probes measure its impact on classification accuracy. Table 3a reports increasing probe gains for GCN*. We repeated the analysis for GAT* under uniform noise (gain in accuracy points): | Dataset | ρ=0\rho=0 | ρ=0.15\rho=0.15 | ρ=0.30\rho=0.30 | ρ=0.45\rho=0.45 | |---|---:|---:|---:|---:| | BlogCatalog | +1.4 | +4.2 | +8.8 | +10.0 | | Flickr | +0.5 | +4.4 | +7.0 | +8.9 | | Roman Empire | +0.6 | +1.7 | +2.4 | +4.0 | | Amazon-Ratings | −0.3 | −0.8 | +0.2 | −0.2 |

    These results confirm the GCN* trends: probe gains increase with noise on three datasets, while remaining negligible on Amazon-Ratings.

  • Overall improvements. JK provides substantial improvements across multiple architectures and datasets, particularly under stronger label corruption. Averaged across noise types and rates, GAT*, GCN*, and GIN improve by +9.3%, +11.0%, and +7.5%, respectively. On BlogCatalog, the average gains reach +20.4% for GAT*, +16.3% for GCN*, and +14.9% for GIN; on Flickr, they reach +13.6%, +24.6%, and +12.1%.

  • Amazon-Ratings and GraphSAGE. These cases help identify when representation preservation provides limited additional benefit.

    • Amazon-Ratings has the lowest recoverability (Γ≈0.10\Gamma \approx 0.10 at every ρ\rho). This is consistent with its negligible probe gain, even though 12.7–14.6% of nodes are classified correctly only from H(0)H^{(0)}. The negative probe gain there shows that node-level complementarity alone is not enough, which is the distinction our analysis is designed to make.
    • GraphSAGE keeps the central-node representation separate from the aggregated neighborhood representation at each layer. This may already preserve part of the node-specific information that JK would otherwise contribute, consistent with its negligible average gain (+0.1%).
  • Noise rate 0. Clean test accuracy at ρ=0\rho = 0 (JK / No JK, mean over three seeds): | Dataset | GAT* | GCN* | GIN | SAGE* | |---|---|---|---|---| | BlogCatalog | 0.972 / 0.933 | 0.974 / 0.960 | 0.956 / 0.795 | 0.967 / 0.971 | | Roman Empire | 0.888 / 0.885 | 0.880 / 0.870 | 0.865 / 0.854 | 0.896 / 0.894 | | Amazon-Ratings | 0.531 / 0.520 | 0.519 / 0.496 | 0.501 / 0.368 | 0.539 / 0.539 | | Flickr | 0.940 / 0.925 | 0.923 / 0.915 | 0.897 / 0.919 | 0.936 / 0.931 |

    For GAT* and GCN*, the relative JK gain is much larger under noise than on clean labels: +4.2% → +20.4% and +1.5% → +16.3% on BlogCatalog, and +1.6% → +13.6% and +0.9% → +24.6% on Flickr. The gain is therefore largely noise-specific. GIN's No-JK variant is already unstable on clean labels (0.368 on Amazon-Ratings), so part of its gain reflects training stability. On Amazon-Ratings, the gain does not grow with noise, which is consistent with its low recoverability (Γ≈0.10\Gamma \approx 0.10). We also added these results to the updated PDF (Table 7).

Reviewer_MUPNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether architecture alone can improve the robustness of GNNs to label noise. It studies a minimal Jumping Knowledge readout that concatenates the pre-propagation node representation with the final propagated representation. Theoretically, the authors show that this readout preserves distinctions that message passing may collapse and contains the standard predictor class. Its optimal linear advantage factorizes into a prediction deficit and the recoverability of that deficit from the pre-propagation representation.

Critical strengths
  1. Instead of changing the training procedure, the paper asks what information the readout should keep, and the modification works with any backbone. The evaluation covers four backbones, four datasets, three noise mechanisms, and three noise rates with corrupted validation labels, and negative cases such as GraphSAGE and Amazon-Ratings are reported openly.

  2. Splitting the advantage into a prediction deficit and its recoverability shows that losing information during propagation is not enough; the lost signal must also be recoverable from the preserved representation.

Critical weaknesses
  1. There are no results at noise rate 0 and no comparisons with an MLP, the full Jumping Knowledge, GCNII or APPNP, or dedicated noisy-label methods. The gains are also concentrated on BlogCatalog and Flickr and rest on only three seeds, and the JK and No-JK GIN models are tuned differently on these two datasets.
Without these, it is difficult to tell whether the gains are specific to label noise or how general they are.


  2. The formal results concern the expressivity of the predictor class and the optimal linear risk with clean targets, and label noise does not appear explicitly in them. The decomposition into deficit and recoverability holds by definition, and the theory does not address why noise enlarges the deficit or how a larger predictor class interacts with noisy labels. This limits how far the theory explains the robustness effect suggested by the title.

Questions for authors
  1. Could you report JK versus No-JK at noise rate 0 for all configurations, so that the gains under noise can be compared with those under clean labels?

  2. Could you add comparisons with an MLP, the full Jumping Knowledge, GCNII or APPNP, and a dedicated noisy-label method for graphs, including that method combined with your readout?

  3. How are the deficit and recoverability computed: on which nodes, and from which trained model? Could the probe analysis be extended to other backbones and noise types, and what accounts for the gap between probe gains and end-to-end gains?

2 comments
AuthorsOpenReview ↗

Response to Reviewer MUPN

  • Noise rate 0: Clean test accuracy at ρ=0\rho=0 (JK / No JK, mean over three seeds): | Dataset | GAT* | GCN* | GIN | SAGE* | |---|---|---|---|---| | BlogCatalog | 0.972 / 0.933 | 0.974 / 0.960 | 0.956 / 0.795 | 0.967 / 0.971 | | Roman Empire | 0.888 / 0.885 | 0.880 / 0.870 | 0.865 / 0.854 | 0.896 / 0.894 | | Amazon-Ratings | 0.531 / 0.520 | 0.519 / 0.496 | 0.501 / 0.368 | 0.539 / 0.539 | | Flickr | 0.940 / 0.925 | 0.923 / 0.915 | 0.897 / 0.919 | 0.936 / 0.931 |

    Results in the updated PDF (Table 7). For GAT* and GCN*, the relative JK gain is much larger under noise than on clean labels: +4.2% → +20.4% and +1.5% → +16.3% on BlogCatalog, and +1.6% → +13.6% and +0.9% → +24.6% on Flickr. The gain is therefore largely noise-specific. On Amazon-Ratings, the gain does not grow with noise, which is consistent with its low recoverability (0.10).

  • Comparison with a dedicated noisy-label method: We compare our readout with GCOD (Wani et al., 2024) on GCN* (BlogCatalog, three seeds): | Noise | | No JK | JK (ours) | GCOD | |---|---|---|---|---| | Uniform | 0.15 | 0.866 | 0.940 | 0.928 ±\pm 0.014 | | Uniform | 0.30 | 0.711 | 0.852 | 0.897 ±\pm 0.012 | | Random | 0.15 | 0.829 | 0.930 | 0.912 ±\pm 0.017 | | Random | 0.30 | 0.698 | 0.862 | 0.891 ±\pm 0.019 |

    JK requires no change to the training procedure and no extra cost. It outperforms GCOD at ρ=0.15\rho=0.15 and recovers most of its gain at ρ=0.30\rho=0.30. Our goal is not to replace dedicated noisy-label methods but to show that robustness can partly emerge from the architecture alone.

  • Comparison with an MLP: We use a linear prediction head to isolate the contribution of the learned representations from the expressiveness of the prediction head. This choice is consistent with our theoretical analysis, which provides an exact characterization of the JK advantage for optimal linear least-squares readouts. We replaced GIN's linear head with a 2-layer MLP and re-ran it on BlogCatalog (uniform noise, ρ=0.15\rho=0.15, three seeds), using identical hyperparameters for JK and No JK: | Head | JK | No JK | |---|---|---| | Linear | 0.919 ± 0.003 | 0.717 ± 0.006 | | MLP | 0.866 ± 0.018 | 0.809 ± 0.011 |

    With the MLP head, the gain is smaller but still clear, so the benefit of JK does not depend on a linear head. We will extend this check to the other noise rates and datasets.

  • Concentration on BlogCatalog/Flickr and GIN tuning.: The concentration is what the analysis predicts. Recoverability is high on BlogCatalog (0.97) and Flickr (0.55) and low on Roman Empire (0.31) and Amazon-Ratings (0.10).

  • The decomposition is an identity. Yes, and that is what makes it useful. Because ΔJK∗=D⋅Γ\Delta^*_{\mathrm{JK}} = D \cdot \Gamma holds exactly, it separates two effects that an "information loss" explanation conflates: losing signal, and being able to recover it. This separation has empirical content. From ρ=0\rho = 0 to ρ=0.45\rho = 0.45, DD grows by 8–72% on every dataset, while Γ\Gamma changes by at most 0.041. The decomposition also explains a case that a naive account cannot: Amazon-Ratings has the largest deficit but the smallest gain, because its recoverability is lowest (Γ≈0.10\Gamma \approx 0.10).

  • Label noise does not appear in the theory. Noise enters through the representations learned at corruption rate ρ\rho, making D(ρ)D(\rho) and Γ(ρ)\Gamma(\rho) noise-dependent. Corollary 1 characterizes precisely when increasing noise leads to a larger JK advantage. Our experiments confirm that this condition holds between ρ=0\rho = 0 and ρ=0.45\rho = 0.45 across all four datasets.

  • The head is fitted to clean targets, not noisy ones. The extension to a head fitted to noisy labels is direct. Under uniform noise with CC classes, the centred noisy target satisfies E[Y~∣Y]=a Y\mathbb{E}[\tilde Y \mid Y] = a\,Y with a=1−ρC/(C−1)>0a = 1 - \rho C/(C-1) > 0. A least-squares head fitted to Y~\tilde Y therefore converges to a PTYa\,P_T Y for the standard readout and to a (PT+Q)Ya\,(P_T + Q)Y for JK. These are the clean-optimal heads rescaled by the same positive constant, so they make the same predictions. The comparison between the two readouts, and hence the role of DD and Γ\Gamma, carries over unchanged. We will add this as a proposition (general class-conditional noise gives P YTP\,Y T for the transition matrix TT, applied identically to both readouts).

  • Deficit, recoverability and probes. For each noise rate and seed, we take the model trained with uniform noise at rate ρ\rho and extract H(0)H^{(0)} (the encoded input features) and H(T)H^{(T)} (the last layer). We then restrict H(0)H^{(0)}, H(T)H^{(T)} and the clean labels to the test nodes, the same nodes used by the probes. Targets are one-hot clean labels. DD is the per-node residual energy of YY after projecting onto col(H(T))\mathrm{col}(H^{(T)}). Γ\Gamma is the fraction of that residual that lies in col((I−PT)H(0))\mathrm{col}\big((I - P_T)H^{(0)}\big). We repeated the full probe analysis on GAT (Figure 2 of the updated manuscript).

Thank you for your response. Authors add experimental results to strengthen the evidence that JK helps under label noise on some datasets and backbones. My main concern on the gap between probe gains and end-to-end gains is unclear. I keep my original rating.

Reviewer_7AHhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work studies the effect of including the encoded input features concatenated to the last layer's representation as input to the classifier or prediction head, in the context of noisy labels at training and validation (but not test) time. They write an analysis of what the jumping knowledge could provide in terms of linear signal, and then ablate on different amount of label noise for several graph architectures.

Critical strengths

I appreciate that the empirical results rely on previous benchmark hyperparameters.

Critical weaknesses

Having corrupted validation and not using weight decay means that the model can overfit, so the selection of the checkpoint may be affecting the results that we observe. In both the empirical and theoretical settings, using only a linear prediction head is very limiting, and additional signal could be extracted with a nonlinear one. But this could increase overfitting, for which it is important to check my first point.

On the analysis part, I disagree with the fact that including the pre-propagation representation adding information for the prediction head being used as explanation for robustness. If using residuals, the information is able to be retained in theory. And in this case, if the signal is already collapsed (like in theorem 1) the question would be why the model has learnt so. I think that most of the theorems are quite simple in what they mean, but are obscured by a lot of notation. Finally, the theory does not match the setup where the head is fitted to corrupted labels, which is what the paper cares about to begin with.

Questions for authors

Can we see training/validation/test curves?

In theorem 1, if the network has learned that two nodes should share a representation, in what cases would they need to be separated? Does this occur in practice, where nodes share a representation that shouldn't?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 7AHh

  • Noisy validation, no weight decay, checkpoint selection: Our experiments are based on a controlled comparison between the same backbone with and without JK. Both variants are evaluated under the same experimental protocol. Therefore, using noisy validation labels and setting weight decay to zero do not inherently favor JK over No JK. Model selection uses noisy validation labels because clean labels are not available in a realistic noisy-label setting. Accuracy measured on noisy labels stays a reliable proxy under class-conditional noise (Chen et al., 2021). We agree that checkpoint selection may still interact with overfitting.

  • Weight decay: It was set to 0 to match the benchmark configurations of Luo et al. (2024), which the reviewer noted as a strength. This choice applies to both JK and No JK, ensuring that neither architecture benefits from a different regularization setting. However, we investigate whether the observed JK gains persist under explicit regularization. We re-ran GCN* on BlogCatalog (uniform noise, three seeds) with weight decay (5×10−45\times10^{-4}), keeping everything else unchanged: | | WD | JK | No JK | |---|---|---|---| | 0.15 | 0 | 0.940 | 0.866 | | 0.15 | 5×10−45 \times 10^{-4} | 0.927 ±\pm 0.009 | 0.887 ±\pm 0.017 | | 0.30 | 0 | 0.852 | 0.711 | | 0.30 | 5×10−45 \times 10^{-4} | 0.861 ±\pm 0.016 | 0.782 ±\pm 0.022 |

    The JK advantage persists and continues to increase with the noise rate. These results indicate that the JK improvements cannot be explained solely by the absence of weight decay. We will extend this analysis to all backbones and datasets.

  • Clean Validation: We re-ran GCN$^*$ on BlogCatalog (uniform noise, weight decay 5×10−45 \times 10^{-4}, three seeds) with clean (oracle) validation labels, used only as a diagnostic: | | Validation | JK | No JK | |---|---|---|---| | 0.15 | noisy | 0.927 ±\pm 0.009 | 0.887 ±\pm 0.017 | | 0.15 | clean | 0.926 ±\pm 0.007 | 0.896 ±\pm 0.004 | | 0.30 | noisy | 0.861 ±\pm 0.016 | 0.782 ±\pm 0.022 | | 0.30 | clean | 0.862 ±\pm 0.015 | 0.791 ±\pm 0.028 |

    Clean selection leaves JK unchanged and improves No JK by about 1 point. The JK gain persists (+3.0 and +7.1) and grows with noise, so it does not come from noisy model selection.

  • Linear head: Refer to response Comparison with an MLP to Reviewer MUPN

  • Residual connections vs. JK. Residual connections and JK operate differently. A residual layer computes H(t+1)=S(t)H(t)+σ(P(t)(H(t);G)),H^{(t+1)} = S^{(t)}H^{(t)} + \sigma(P^{(t)}(H^{(t)};G)), where the skip branch (S(t)H(t)S^{(t)}H^{(t)}) is combined with the propagated representation at each layer. Importantly, GCN*, GAT*, and SAGE* already employ learned residual connections. The JK gains observed for GCN* and GAT* demonstrate that direct representation preservation can provide benefits beyond residual connections.

  • Interpretation of Theorem 1. Theorem 1 establishes a sufficient condition under which JK strictly enlarges the predictor class, independently of the propagation operator. It does not assume that information collapse necessarily occurs, nor does it aim to explain why training produces particular representations. Rather, it formalizes the structural distinction between the two readouts. Proposition 1 then provides the more informative result: an exact decomposition of the optimal linear JK advantage into prediction deficit and recoverability. This identifies when the additional representation contributes useful predictive information, rather than increasing the predictor class.

  • Theory vs. noisy training: In our analysis, the representations H(0),H(T)H^{(0)},H^{(T)} are those learned at noise rate , so noise enters through them. The risks are then measured against clean targets. This gives the best clean accuracy any linear head can reach on those representations, not the accuracy of a head fitted to noisy labels.

  • When should nodes with the same representation be separated?: Nodes should remain distinguishable when they have different clean labels but propagation makes their representations indistinguishable to the prediction head. While Theorem 1 considers exact collapse, our experiments examine the more general loss of accessible predictive information. Figure 1 shows increasing node-level complementarity under noise, while Table 3b shows increasing prediction deficit with relatively stable recoverability. Proposition 1 does not require exact representation collapse, but characterizes when predictive information remains recoverable from H(0)H^{(0)}. They should be separated when they share an aggregated neighborhood representation but have different clean labels. This is common in heterophilous graphs, and label noise makes it worse. It does happen in practice. The linear-probe analysis (Figure 1) also shows that the share of nodes classified correctly only by the H(0)H^{(0)} probe grows with the noise rate.

  • Can we see training/validation/test curves?: Figure 4 from the updated manuscript shows them.

  • On "using noisy validation labels and setting weight decay to zero do not inherently favor JK over No JK", I think they can. For instance, if both models learn the clean signal at a similar rate, but JK takes longer than no-JK to fit the corrupted training labels, then cutting training early can favor JK. This is where I see a connection with memorization theory which I think this paper needs. Even a linear classifier can first learn the clean signal and later memorize the corrupted labels, see e.g. Liu et al. (2020).
  • Thank you for the additional experiments. Interestingly, the difference between JK and no-JK seems to be smaller with weight decay. At lower noise wd improves no-JK and slightly lowers JK performance. Similarly, no-JK gets better with clean validation, while JK changes very little. I see a similar pattern with the MLP, where no-JK gets better but JK can even get worse. I do not see how the current theory accounts for this.
  • I do not argue that the gap doesn't exist, but my point is about whether the theory captures why it occurs. I understand that the representations are learned under noisy supervision but the linear head is then fitted to clean labels. But in the experiments, being able to separate two nodes can also help the head memorize a (wrong) flipped label. What I am missing is how much of the JK advantage can be explained by this additional clean signal when the head has to learn from corrupted labels.
  • The training validation curves in Figure 4 are very difficult to read, several lines are on top of each other, and I cannot tell if it's JK or no-JK. This is not helping answer my original question about the mechanism at play. As an example of what I am looking for, on the training nodes whose labels were flipped, it could be the case that no-JK starts predicting the corrupted labels earlier than JK.
  • Early stopping and slower memorization. The reviewer proposes that JK might fit corrupted labels more slowly than No JK, so that stopping early would favor JK. This is a possible mechanism, but to our knowledge no existing result shows that JK memorizes noisy labels at a different rate than No JK. The early-learning pattern (Liu et al., 2020) is general and does not imply a difference between the two architectures, so we treat it as a conjecture.

    Our clean-validation experiment already controls for this channel. With oracle clean validation, each model is stopped at its own best clean epoch, so a difference in memorization speed cannot favor either model through the stopping rule. The JK gain remains (+3.0 at noise rate 0.15, +7.1 at 0.30).

  • Smaller gap with weight decay, clean validation and the MLP head. The reviewer is right that the gap shrinks. At noise rate 0.15, weight decay reduces it from 7.4 to 4.0 points, and at 0.30 from 14.1 to 7.9 points. Clean validation reduces it to 3.0 and 7.1 points. No JK benefits more than JK from regularization and clean selection. Part of the No JK deficit in our original protocol therefore came from missing regularization and noisy selection. We say so in the revision and add standard deviations to the weight-decay-0 rows. The remaining gap is positive in every setting and grows with the noise rate.

    We agree that our theory does not predict this pattern. Theorem 1 and Proposition 1 describe what the learned representations make available to a linear head, not how regularization or selection affect training. We state this scope explicitly and no longer claim that the theory explains the size of the gap.

  • Additional clean signal versus memorization by the head. The reviewer is right that separating two nodes can also help a head fit a wrong, flipped label. Two observations from our results bear on this:

    • All accuracies we report are measured against clean test labels. Memorizing a flipped training label cannot increase clean test accuracy, so it cannot account for a test-accuracy gain.
    • JK's head takes a wider input (2d2d instead of dd) and so has more capacity to memorize. If memorization were the main driver, JK should be hurt more as noise grows. Its advantage instead grows with the noise rate.
  • Figure 4. We apologize that the curves were hard to read. Our aim was to give the reviewer as much information as possible.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

The discussion period is coming to a close, and we have not yet received any follow-up on our rebuttal. We have tried to address every concern raised in the reviews, including additional experiments and clarifications where requested.

We would appreciate it if you could let us know whether our responses resolve your concerns, or whether anything still needs clarification. We are happy to provide further details before the deadline.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.