acceptodds
Under review as a conference paper at ICLR 2027

Neutralizing 0th-Order Residual Interferences in Layer-wise Relevance Propagation via Geometric Trajectory Realignment

Abstract

Standard Layer-wise Relevance Propagation (LRP) operates under the premise that 0th-order residual terms, even if not strictly eliminated, do not affect the essential properties of relevance propagation. However, this study reveals that when introducing a simple constant 0th-order term as a minimal 0th-order probe, the linearity of the distribution persists under constrained input methods, causing noise to accumulate during relevance propagation. This phenomenon induces a deviation () within the - parameter space and misaligns geometric trajectories, thereby structurally violating LRP's conservation principle and ultimately leading to fatal signal collapse in deep layers. To empirically validate this theoretical distortion mechanism, we introduce the -rule—a normalization mechanism that neutralizes the linear effects of the 0th-order terms by aligning geometric trajectory slopes to unity. Our analysis of CNN architectures (e.g., VGG, ResNet, DenseNet) confirms the validity of our hypothesis: the -rule maintains a constant variance of , effectively suppressing distribution-induced noise and improving explanation fidelity. Conversely, the marginal improvements observed in Vision Transformers (ViTs) expose the limitations of the simplified constant 0th-order assumption in complex non-linear models. Despite this boundary of validity, our empirical study unequivocally proves that the distribution of 0th-order terms alone generates critical noise in LRP, establishing an essential theoretical foundation for designing novel rules to address non-linear residual dynamics in future Transformer architectures.

Primary area:
interpretability and explainable AI
Keywords:
Explainable AI (XAI), Layer-wise Relevance Propagation (LRP), Signal Collapse, 0th-Order Residual Terms, Geometric Trajectory Alignment

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

The paper argues that the treatment of 0th-order (bias or reference-point) residuals in standard LRP causes signal collapse and sign flips in deep networks. A constant 'minimal probe' b^m (Assumption 3.1) is inserted into an αβ-style rule (Eq. 5), the surviving term (Eq. 7) is interpreted as an angular deviation θ (Def. 3.2), and the proposed π-rule (Eq. 15 & 16) is claimed to help CNNs and ViTs.

Critical strengths

Whether the handling of bias terms contributes to relevance distortion in deep networks is a fair question, and isolating it with Uniform LRP (§5.1, Appendix B) is a good idea.

Critical weaknesses

I do not think the claim that 0th-order terms cause signal collapse, and that the π-rule removes the cause, is established, for four reasons.

  1. The 'noise' of Eqs. 6 & 7 does not seem to me to follow from the αβ-rule or DTD. Eq. 5 adds b^m to every (i,j) term, although Σ_i Z_ji R_j = R_j - c·b̃_j (Eqs. 2 & 3) already covers the bias contribution. The αβ-rule also conserves relevance only for α - β = 1, yet the sweep (§5.2), the 1:1 upper-layer rule of LRP_com and Figures 3 & 4 lie outside it (α - β = 0 at 1:1), so sign flips are expected without any b.

  2. §4.3 itself ends with R_bias ∝ (α − β)b^m, which is non-zero unless α = β. With π = 2PM/(P+M), the coefficient becomes π(α − β)/(α + β), i.e., -10 for the αβ-rule against +14 for the π-rule at α:β = 2:1, P = 30, M = 70. Moreover, α_π/β_π = (α/β)(M/P), so the rule does not 'propagate the originally intended α:β ratio' (§4.2).

  3. For ViTs the abstract reports 'marginal improvements', yet LRP_com^π is worse than LRP_com^αβ on all six metrics for ViT-B and ViT-L (12 of 12 comparisons, Tables 2-7). Plain LRP_(ε) beats LRP_com^π in 17 of 17 CNN rows of Table 2, so the 'consistently superior' claim (§5.3) is not particularly strong (only a weak baseline is compared). SGLRP is omitted from Table 1, and in Table 2 the π-rule is worse in 7 of 17 SGLRP rows.

  4. In Fig. 2, θ depends only on P^m and M^m and is zero for the π-rule by construction, so its suppressed variance is close to tautological. In Fig. 4, rescaling (α, β) jointly multiplies both terms of Eq. 6 equally, so Eq. 7 predicts a scale-invariant noise-to-signal ratio. Given that the magnitudes could very well overflow after ~8 layers, this should also be ruled out.

I also found there to be some significant overstatements and information I could not find:

  1. The "unequivocally proves" claim in the abstract very much overstates the unchecked assumption, and I could not find the derivations cited in the Reproducibility Statement in Appendices A or B.
Questions for authors
  1. How is θ computed for the π-rule in Figure 2, given θ = 0 by construction?

  2. Do collapse and sign flips persist at α - β = 1 (e.g., 2:1) once biases are zeroed, and how do Figures 3(c) and 4 look in float64?

  3. How does an αβ-rule with the π-rule's effective ratio (αM:βP) perform on its own?

  4. Please reconcile 'marginal improvements' for ViTs with the 12 of 12 worse comparisons, and add SGLRP to Table 1 with paired confidence intervals.

3 comments

Dear Reviewer mrLk,

We sincerely thank you for your deep and highly critical examination of both our mathematical premises and experimental details. Your sharp feedback has been invaluable in helping us identify and correct overstated claims and omitted prerequisite conditions. We address your critical concerns below.

Q1 & Q2 We apologize for not explicitly detailing the normalization prerequisites of the Z-rule in the main text, which understandably led to confusion regarding Equation (5) and the noise propagation dynamics. However, we must respectfully point out that the standard LRP αβ\alpha\beta-rule operates under the condition of α\alpha + β\beta = 1, not α\alpha - β\beta = 1 as you suggested. Furthermore, our implementation strictly adheres to the standard prerequisite where the sign of the inhibitor is explicitly normalized in the Z-rule denominator (i.e., z−→−Z=−z−/∑z−z^- \rightarrow -Z = -z^- / \sum z^-). Therefore, the expectation of sign flips based on the α\alpha - β\beta = 1 assumption is rooted in a misaligned mathematical premise. Regarding Figure 2, θ\theta is not tautologically forced to zero by construction. The θ\theta values are empirical observations derived via Equation (9), based on the actual activation counts (P and M) during the forward pass. The core structural finding of our paper is not that we mathematically force θ\theta to converge to zero. Rather, it is that by aligning the trajectory via the π\pi-rule, we successfully suppress the variance of the accumulated noise, thereby empirically demonstrating that the previously ignored 0th-order bias term exerts a fatal, destructive impact on the actual propagation process.

Q3 We completely agree with your observation that the initial ratio of the ϵ\epsilon-rule demonstrates strong performance in specific configurations (Table 2). However, the primary contribution of the π\pi-rule is not to marginally outscore every baseline on every metric, but to provide fundamental robustness against heuristic parameter tuning. While the standard αβ\alpha\beta-rule requires exhaustive searches to find optimal ratios, the π\pi-rule fundamentally neutralizes structural noise, significantly reducing performance fluctuations across different parameter settings. We gladly accept your constructive suggestion and will add a comparative experiment applying the π\pi-rule's effective ratio (αM:βP\alpha M:\beta P) to the standard αβ\alpha\beta-rule in our revision, which will further highlight this robustness.

Q4 We fully concede your point here. Characterizing the ViT results as "marginal improvements" was an inappropriate overstatement. As you correctly noted, the performance declined across all 12 comparisons, representing a clear limitation of our current methodology. As briefly discussed in our "Boundary of Validity" section, the minimal probe assumption is insufficient to capture the highly non-linear residual dynamics (such as GELU) present in Transformer architectures. We humbly accept your critique: we will immediately remove the overstated claims from the Abstract and explicitly document the performance degradation in ViTs as a core limitation. Furthermore, we will add the SGLRP results to Table 1 with paired confidence intervals, exactly as requested.

We are deeply grateful for your rigorous review, which will unequivocally improve the honesty, precision, and soundness of our final manuscript.

Do your experiments implement R = α·(positive share) − β·(negative share) with β ≥ 0, as in Eqs. 5 and 16, or a signed β? In the first case α + β = 1 gives total relevance α − β (zero at 0.5:0.5), and in the second Eqs. 5 and 16 need rewriting.

Dear Reviewer mrLk,

Thank you for your prompt follow-up.

To answer your question directly: our experiments implement β≥0\beta \ge 0, exactly as written in Equations 5 and 16, and the equations do not need to be rewritten. The use of the negative sign before β\beta is not a peculiar notation introduced by us, but rather strictly follows the explicit normalization steps of the standard LRP formulation (e.g., as referenced in "Layer-wise relevance propagation: an overview" by G. Montavon et al.).

To clear up any confusion, here is the step-by-step derivation from the most basic LRP formula to our Equation 5, explicitly showing how the negative sign emerges during normalization.

1. The Basic LRP Formula The fundamental propagation of relevance can be expressed as: R=∑ZRR = \sum Z R where ZZ is the contribution score and RR is the relevance score.

2. Conceptual αβ\alpha\beta-rule When separating the positive and negative contributions, the rule is commonly conceptualized in literature as: R=α∑Z+R+β∑Z−RR = \alpha \sum Z^+ R + \beta \sum Z^- R (This is the shorthand form that most authors and reviewers are familiar with.)

3. Explicit Normalization of Contributions (ZZ-rule) However, to rigorously analyze the structural interaction of our "minimal probe," we cannot rely on the abbreviated form. We must explicitly unroll how the raw contributions (zz) are normalized.

Let zz be the raw contribution. The normalized positive and negative contribution scores (Z+Z^+ and Z−Z^-) are defined by explicitly factoring out the negative sign of the inhibitor to represent the "share" as a positive fraction: Z+=z+∑z+Z^+ = \frac{z^+}{\sum z^+} Z−=−z−∑z−Z^- = -\frac{z^-}{\sum z^-}

4. Final Expanded Formulation By substituting these explicitly signed normalizations back into the conceptual αβ\alpha\beta equation, we get the fully expanded formulation: R=α∑(z+∑z+)R−β∑(z−∑z−)RR = \alpha \sum \left( \frac{z^+}{\sum z^+} \right) R - \beta \sum \left( \frac{z^-}{\sum z^-} \right) R

This fully expanded form precisely matches the "−β-\beta" formulation presented in our Equation 5. Because the negative sign is mathematically extracted during the Z−Z^- normalization step, applying the condition α+β=1\alpha + \beta = 1 (with β≥0\beta \ge 0) does not violate total relevance conservation.

We fully concede that due to space constraints, our brief definition of ZZ (Z=Jx/f(x)Z = Jx/f(x)) right before Equation 5 was too condensed and omitted this crucial step-by-step sign separation, making it difficult to bridge the gap from the basic formula. We will explicitly add this full derivation in the revised manuscript to ensure the mathematical rigor is clear to all readers.

Thank you again for your careful reading and for helping us identify this missing theoretical bridge.

Reviewer_FxsjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether zeroth-order residuals, in conjunction with unequal populations of positive and negative contributions, meaningfully distort layer-wise relevance propagation (LRP). The authors propose the π\pi-rule, a normalization of α\alpha and β\beta based on the harmonic mean of these contribution counts. They interpret this normalization as geometric trajectory alignment and claim that it restores conservation and prevents signal collapse. Their theoretical analysis uses a constant-residual approximation and a random-walk model with independent increments. Empirically, the authors evaluate the method on traditional ImageNet using VGG, ResNet, DenseNet, and ViT models across uniform and composite LRP configurations, with different propagation settings and six perturbation metrics.

Critical strengths
  1. In addition to the theoretical motivation, I found the consistent improvements across all tested CNN configurations encouraging (Tables 2–7). The proposed π\pi-composite method improves all six reported perturbation metrics relative to the authors' αβ\alpha\beta-composite baseline. Although these tests are relatively simple, measuring how predictions change when features are perturbed according to their attributed importance showed consistent gains, suggesting that the proposed reweighting may produce more useful feature rankings for these models.

  2. I also appreciated the simplicity of the π\pi-rule, which adjusts the weights of positive and negative contributions according to the sizes of the two groups, using a shared normalization factor based on their harmonic mean. This gives the proposed adjustment a compact mathematical form that is straightforward to inspect and analyze. The less favorable results on vision transformers (ViTs) are also informative, since they help identify the method's current limitations within the authors' own evaluation.

Critical weaknesses
  1. Conservation doesn't seem to be generally preserved

My main concern is the claim in Section 4.2 that the proposed normalization preserves relevance. In LRP, conservation means that the sum of the relevance scores, including both positive and negative values, remains unchanged as they pass backwards through a layer. For the standard αβ\alpha\beta rule, the usual condition is α−β=1\alpha-\beta=1, assuming that there is no bias or numerical stabilizer and that both contribution groups are nonempty and normalized separately. The rescaling in Eq. (15) does not generally preserve this condition.

For example, if we consider f(x)=2x1−x2f(x)=2x_1-x_2 at x=(1,1)x=(1,1), with a relevance score of R=1R=1 to redistribute and (α,β)=(2,1)(\alpha,\beta)=(2,1). The standard rule assigns (2,−1)(2,-1), whose entries sum to 11. With one counted entry in each group, Eq. (15) changes the coefficients to (2/3,1/3)(2/3,1/3). When applied as in the supplied implementation, this gives relevance scores (2/3,−1/3)(2/3,-1/3), whose sum is only 1/31/3. A small test of the supplied code reproduces this discrepancy up to numerical tolerance. Because this example contains no bias, the failure cannot be explained by the residual term that the method is intended to address.

I also found the connection between the theoretical model and the implemented rule insufficiently clear. Eq. (5) normalizes contributions by the neuron's total activation, while the implementation normalizes the positive and negative groups separately. The paper therefore needs to explain how the analysis applies to the rule being tested. Similarly, although the mean and variance calculations in Eqs. (11)–(14) are correct for the assumed independent random walk, the paper does not establish that this process describes the actual propagation dynamics. Increasing variance, by itself, also does not explain why every entry of an explanation would become zero.

  1. Collapse mechanism and its measurement are not clear

I am also concerned about the difference between the stated definition of collapse and the supplied measurement code. Appendix B.1 defines collapse as an explanation whose entries all vanish, whereas anl_LRPab_ratio.py, line 42, uses (vec_lrp==0).sum(1).bool().sum(), which counts a map whenever any entry is zero. Thus, a map such as [0,1,2,3][0,1,2,3], which still contains several nonzero scores, would be counted as collapsed. If this script produced Figure 3(c), the collapse frequencies and associated averages need to be recomputed using the stated definition.

The implementation also provides another possible explanation for disappearing relevance. The Add operator used for skip connections rescales relevance using its incoming signed total, so a nonzero vector such as (1,−1)(1,-1) can be erased because its entries sum to zero. This behavior is reproduced in a small example without bias and is particularly relevant when α=β\alpha=\beta, since separately normalized positive and negative contributions can already cancel in their total. In addition, the π\pi counts are computed across the entire batch, which means that an image's relative feature scores can change when unrelated images are processed alongside it. Clarifying which code produced the reported results, and separating these effects from those of residual terms, is necessary to assess the proposed explanation of collapse.

Finally, I would suggest a more narrow interpretation of the empirical gains than the one presented. Although π\pi-composite consistently improves the authors' αβ\alpha\beta-composite baseline on CNNs, the baseline labelled ϵ\epsilon performs better on every CNN row for most-relevant-first removal (MoRF), positive perturbation and deletion. On both tested ViTs, π\pi-composite performs worse on all six metrics than αβ\alpha\beta-composite. These comparisons qualify the scope of the reported improvements. Moreover, the perturbation scores, which assess how feature rankings relate to changes in predictions, do not in my opinion directly verify conservation nor establish the cause of the collapse.

Questions for authors

I found the consistent improvements over the reported αβ\alpha\beta-composite baseline on CNNs in Tables 2–7 encouraging, and I would welcome clarification on the points below. I am open to revising my assessment and score in light of your responses and any supporting evidence.

  1. Could you clarify what is meant by conservation in Section 4.2, and under which assumptions the π\pi-rule satisfies it? In the simple linear example, the supplied implementation receives an incoming relevance score of 11 and returns scores whose sum is approximately 1/31/3. If this example falls outside the intended setting, or if an additional normalization step is required, it would be helpful to make that explicit. I would also appreciate clarification of how the contribution scores defined just before Eq. (5) relate to the separate normalization of the two branches in the code.

  2. Could you confirm how explanations were counted as collapsed in Figure 3(c)? The expression on line 42 of anl_LRPab_ratio.py counts a map if it contains any zero entry, whereas Appendix B.1 describes collapse as the entire explanation map vanishing. If the figure was generated with a different implementation, sharing the relevant code would help resolve this concern. If the supplied script was used, could you report whether applying the stated definition changes the collapse frequencies and associated averages?

  3. Could you clarify whether the reported experiments used contribution counts pooled across images and the supplied Add rescaling at skip connections? If so, it would be helpful to know whether the CNN gains persist with counts computed separately for each image, and whether the Add rescaling contributes to collapse by erasing nonzero scores whose total is zero. A focused sensitivity check using α−β=1\alpha-\beta=1, such as (α,β)=(2,1)(\alpha,\beta)=(2,1), would also help assess how much the findings depend on the equal-coefficient setting described in Appendix A.3.

  4. Could you explain how the random-walk model in Section 3.4, Eqs. (10)–(14), relates to the actual propagation process? The mean and variance formulas hold for the stated scalar model, whose independent steps take values +Nb+Nb or −Nb-Nb with fixed probabilities, with Eqs. (13)–(14) additionally assuming p=q=1/2p=q=1/2. What remains unclear to me is why increasing variance should cause every score in an explanation to vanish. A derivation or controlled example connecting these steps, while accounting for the implementation effects above, would help me assess the proposed mechanism.

1 comment

Dear Reviewer Fxsj,

We sincerely thank you for appreciating the mathematical simplicity of our \pi-rule and the consistent empirical improvements it yields across CNN architectures. We also deeply appreciate your rigorous inspection of our formulation and implementation, which has helped us identify areas where our explicit assumptions were not clearly communicated. We address your specific concerns below.

Q1 We appreciate your careful derivation of the toy example and completely understand your concern regarding the mathematical discrepancy. This confusion arose because we inadvertently omitted the explicit prerequisite conditions of our implementation in the text. Our π\pi-rule operates strictly under the standard prerequisite condition of α\alpha + β\beta = 1. Furthermore, following the standard Z-rule practices, the sign of the inhibitor is explicitly normalized in our codebase (i.e., z−→−Z=−z−/∑z−z^- \rightarrow -Z = -z^- / \sum z^-). When this explicit normalization condition is met within the α\alpha+β\beta=1 framework, our proposed formulation strictly satisfies the global conservation rule of LRP. We will explicitly add these foundational assumptions and normalization steps to Section 4.2 to resolve this ambiguity.

Q2 You are absolutely correct that there is a discrepancy between our textual description ("entire explanation map vanishing") and the actual measurement code used ((vec_lrp==0)...). We fully concede that the term "zero-valued explanation map" was an overstatement. However, the core phenomenon we aimed to measure and highlight is "extreme sparsity." As structural noise explodes, the activation values become so sparse that they converge to floating-point zero, causing the map to visually and structurally collapse. Because the exact same measurement criteria were applied fairly and uniformly to both the standard αβ\alpha\beta-rule and our π\pi-rule, the relative performance gap and the effectiveness of the π\pi-rule remain highly valid. Following your feedback, we will correct the manuscript to accurately describe this phenomenon as "collapse due to extreme sparsity" rather than using the absolute "zero-valued" terminology.

Q3 We thank you for pointing out the potential nuances regarding batch processing and the Add operator (skip connections). To ensure reliability and reproducibility, our codebase strictly adopts widely validated standard LRP implementations. The application of the π\pi factor is rigorously targeted at aligning the geometric trajectories of Linear and CNN layers. The mathematical cancellation concerns you raised regarding the Add operator are effectively controlled under the explicit normalization conditions clarified in our response to Q1. Therefore, the core empirical finding of our study—that the π\pi-rule consistently improves explanation fidelity across CNN structures—remains robust.

Q4 We agree that the theoretical random-walk model abstracts certain complex dynamics of the actual deep propagation process. However, the primary purpose of this mathematical modeling is to formally demonstrate that the 0th-order bias is not a benign constant, but an active generator of structural noise. The explosive increase in this theoretical noise level (variance) directly translates into the two fatal empirical distortions we present: (1) Severe Sign Flips and (2) Signal Collapse (extreme sparsity). The unconstrained variance amplification causes the relevance scores to aggressively fluctuate and cancel each other out during propagation, which is the root mechanism driving the signals to vanish in deep layers. We will expand our discussion in the text to explicitly bridge the theoretical variance amplification with these two empirical failure modes.

We are grateful for your constructive critique, which will significantly improve the precision and clarity of our final manuscript.

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

The paper investigates interference from zeroth-order residuals in layer-wise relevance propagation (LRP). By approximating these residuals as constants within each layer, it argues that an imbalance between the sets of positive and negative contributions causes geometric trajectory shifts and uses a random-walk model to explain signal collapse during propagation through deep layers. Based on this analysis, the authors propose the π-rule, which uses set sizes and harmonic-mean normalization, and evaluate it on all 50,000 ImageNet validation images across multiple CNN and ViT architectures. The results show that the method substantially improves composite LRP configurations for some CNNs, while gains under other configurations are limited and some metrics decline on ViTs.

Critical strengths
  1. The problem has practical significance. It focuses on sign flips, relevance attenuation, and all-zero explanation maps in deep-layer explanations, addressing the reliability of explanation methods themselves.
  2. Linking the sizes of positive and negative contribution sets to propagation bias provides a concrete starting point for analyzing instability in LRP.
  3. The π-rule is implemented by adjusting propagation coefficients without retraining the model, making it easy to integrate and conduct comparative experiments.
Critical weaknesses
  1. Section 3.4 directly assumes that noise across layers is independent and identically distributed, without deriving this assumption from the actual propagation recurrence, and ignores the role of inter-layer propagation operators. Even under these assumptions, Equation (14) only shows that the variance of accumulated noise increases; it does not imply that the relevance signal converges to zero.
  2. A deterministic algorithm and evaluation on the full validation set do not eliminate uncertainty about generalization to the population and therefore do not justify the claim that statistical analysis is unnecessary.
Questions for authors
  1. Can Equation (5) be derived from the standard propagation rules, and can a complete proof of conservation be provided for Equation (16)?
  2. How do signal collapse and the gains from the π-rule change when network biases are removed, the residual magnitude is adjusted, or only scale normalization is applied?
1 comment

Dear Reviewer dZDc,

We sincerely thank you for recognizing the practical significance of our problem and the experimental utility of the π\pi-rule. We appreciate your rigorous and constructive feedback, which has helped us clarify our claims. We address your specific concerns below.

Q1 We completely agree with your rigorous mathematical observation. As you correctly pointed out, Equation (14) demonstrates that the variance of the accumulated noise increases, rather than mathematically proving that the relevance signal inherently converges to zero. We will correct the phrasing in our manuscript to ensure this distinction is clear. However, we would like to gently remind the reviewer that this uncontrolled amplification of variance is precisely the core theoretical mechanism we aim to highlight. When the minimal 0th-order probe couples with the network topology, it generates linear noise (variance) that persistently accumulates. In deep layers, this amplified variance completely overwhelms the pure 1st-order relevance signal. This structural destruction is what empirically manifests as the "signal collapse" (e.g., zero-valued explanation maps) observed in our experiments. We will carefully revise Section 3.4 to reframe this phenomenon from "convergence to zero" to "the pure signal being overwhelmed by exploding variance."

Q2 We fully concede your point that evaluating a full validation set does not eliminate uncertainty regarding generalization to the broader population. Our original statement claiming that statistical validation is unnecessary was poorly phrased and unintentionally misleading. Our intent was solely to emphasize that because LRP is a deterministic algorithm, evaluating the entirety of the 50,000 ImageNet validation images completely precludes "statistical sampling bias" within that specific benchmark. Since defining a proper probabilistic sample space for a deterministic algorithm on a fixed benchmark is highly ambiguous, we opted for a full-set evaluation. We will tone down the claim in the revised manuscript (specifically in Appendix A.1) to state only that we "eliminated sampling bias on the validation set", removing any broader claims against statistical analysis.

Thank you again for your valuable insights, which will undoubtedly strengthen the clarity and rigor of our final manuscript.

Comments on the paper

1 comment
Area_Chair_vBrpOpenReview ↗

Reviewer's responses to authors

Dear reviewers,

If not done already, could you please take a look and reply briefly to the "Official Comment(s) by Authors" / Rebuttals, possibly by an acknowledgement that you have seen it.

Feel free to respond in more detail, in particular, if you have further questions or opinions to the replies from the authors.

With best regards, an AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.