acceptodds
Under review as a conference paper at ICLR 2027

ViTViz: Analyzing Attention Redistribution Under Adversarial Attacks in Vision Transformers

Abstract

Vision Transformers (ViTs) have achieved state-of-the-art performance in image classification, yet they remain vulnerable to adversarial attacks. Since attention is a fundamental mechanism underlying these models, understanding how it behaves under adversarial perturbations is crucial. However, to the best of our knowledge, the literature lacks tools for systematically exploring attention under adversarial regimes. To address this gap, we introduce ViTViz, an open-source tool for exploring attention under adversarial attacks. ViTViz enables researchers to independently investigate key dimensions of attention—including the aggregation operator, anchoring class, layer, attention head, attack step, and perturbation budget—on any plain ViT classifier with a single class token, in the timm or the Hugging Face layout. For the benchmark's six models, it places each measurement at its percentile among images of the same model, attack, operator, and outcome. Furthermore, we introduce an attention redistribution analysis based on three measures, each answering a different question: how far attention mass moved, whether the distribution changed at all, and whether attention spread or concentrated. As a proof of concept, we applied this analysis to 1,000 ImageNet images subjected to five attacks, using six ViT variants and two attention aggregation operators, one class-agnostic and one anchored on a class label. On the same images, the two operators disagree about whether a successful attack moves attention further than a failed one. How far attention moved therefore says something about an attack only when the operator and anchor that produced the measurement are reported.

Primary area:
interpretability and explainable AI
Keywords:
Vision Transformers, explainability, attention-based explanations, attention rollout, adversarial attacks, interactive visualization tool, evaluation of explanations

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_JT71OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ViTViz to study how ViT attention changes under adversarial attacks. By comparing Attention Rollout and GAE, the authors show that different ways of reading attention can lead to very different, and sometimes opposite, conclusions about the relationship between attack success and attention displacement.

Critical strengths

The experiments are relatively systematic, covering multiple ViT models, attacks, and data domains. The authors also use bootstrap and several control settings involving perturbation budget, anchor, and normalization to validate the main observations. In particular, Rollout and GAE show a significant directional reversal on GTSRB, indicating that the choice of attention reading can indeed affect the conclusion.

Critical weaknesses

1.The main contribution is mostly a combination of existing Rollout, GAE, standard redistribution metrics, and a visualization tool. There is no new attack or attention method. Since Rollout is class-agnostic while GAE uses class gradients, some disagreement is not very surprising. To me, this is closer to an engineering/measurement study than a strong methodological contribution.

2.Rollout and GAE differ in several ways at the same time, including class anchor, gradient weighting, aggregation, and normalization. So it is hard to tell what actually causes the disagreement. The current controls only explain part of it.

3.Personally, PGD and MIM are too successful in some settings, leaving very few unflipped samples. This makes the successful-vs-failed comparison less convincing in those cells. For example, comparing 990 flipped samples with only 10 unflipped samples is clearly imbalanced. The GTSRB experiment also changes the dataset, checkpoint, and class space together, so I do not think the effect can be attributed only to domain shift.

4.The experiments show that different attention readings can give different conclusions, but they do not really explain why. They also do not establish a clear mechanism between attention redistribution and attack success. The paper’s own variance analysis even shows that image identity matters much more than attack identity.

5.The ablation is still not complete. I would like to see a more direct factor-by-factor analysis of the differences between Rollout and GAE. Also, only comparing two attention operators is not enough to support a broad claim about attention reading in general.

6.The attack set is also somewhat outdated. Recent ViT-specific attacks such as ATT (NeurIPS 2024) and FPR (CVPR 2025) are not included. It is unclear whether the same conclusions still hold for stronger and more recent attacks.

Questions for authors

See weaknesses.

2 comments

We thank Reviewer JT71 for the thoughtful review and for highlighting the systematic experiments across models, attacks, and data domains.

W1 (Contribution and class information). We benchmark displacement by model, attack, operator, and outcome and test which findings carry across readings. With a random anchor, GAE and LeGrad retain 8 and 6 of 22 same-direction separations, and GMAR keeps all 15 of its original separations. Using the predicted class moves GAE’s map by 64%–89% of the mean attack displacement and LeGrad’s by 1.7–2.2 times. Random anchoring changes GMAR’s displacement by a median of 1.1% per image. Aggregation drives its eight-cell gap from rollout.

Changes in the revision. Table 11 reports random anchors, Appendix I predicted-class anchors, and Appendix P GMAR’s decomposition.

W2 (Separating factors). We varied operator choices on the same images. Across-layer gradient weighting has the largest Shapley share for all three measures. On raw attention, GAE’s composition gives rollout up to a constant. Shares are in cells that separate, for the original five attacks.

FactorW1‾\overline{W_1}JSD∣ΔH‾∣\vert\overline{\Delta H}\vert
Gradient across layers+6.5+5.5+8.5
Gradient within a layer+2.5+2.0-2.3
Composition+2.5+0.5-0.3
Higher-order terms+1.00.0-1.0
Normalization-0.5-2.00.0
Gap8 → 2011 → 177 → 12

Changes in the revision. Section 5 and Table 18 (Appendix P) give the decomposition. Table 11 controls for the anchor.

W3 (Small samples and GTSRB). The 20-unflipped-image floor excludes splits such as 990/10. We replaced the percentile bootstrap with a Welch-statistic permutation test. With 20 unflipped images, false separation fell from 8.5% to 5% in simulation. Simulated power was 38%, so we report each cell’s resolution. The original rollout and GAE verdicts are unchanged. We also tested five more GTSRB checkpoints. Rollout and GAE agree in 20 of 26 estimable cells, including 11 of 16 under FGSM, PGD, MIM, and TGR. The PGD reversal did not recur in the three with estimable PGD results. Dataset and class space still change together in the ImageNet–GTSRB comparison.

Changes in the revision. Section 5 and Appendix N describe the test. Section 6.3 and Appendix K report the checkpoints.

W4 (Mechanism and image identity). We localized the contrast to late layers and patched their attention. Restoring the clean pattern recovers 12%–39% of the lost margin and undoes 2%–14% of flips. Restoring the full attention output recovers 75%–94%. In the reverse test, another clean image’s pattern lowers the true-class probability at least as much as the attacked pattern in five of six cells. Late attention plays a partial role. Most flips survive restoration.

With the original five attacks under GAE, image identity explains 42%–63% of displacement variance and attack identity 3%–24%. Across rollout and GAE, the image–operator interaction explains 17% and the attack 7%. This motivates comparing flipped and unflipped images within each model and attack and using the run’s own reference group.

Changes in the revision. Appendices Q and R add the depth analysis and patching. Section 6 and Appendix B give the variance estimates.

W5 (Four operators). We added LeGrad and GMAR and decomposed three operator gaps. Rollout, GAE, LeGrad, and GMAR separate 11, 29, 35, and 22 of 35 estimable cells. On the original five attacks, no pair agrees in more than 16 of 24 cells. Across-layer gradient weighting, logit choice, and aggregation drive different pairs. We recommend comparing multiple readings and reporting where their conclusions agree.

Changes in the revision. Table 3 and Appendix J extend the comparison. Appendix P gives the three decompositions.

W6 (Recent attacks). We added FPR and ATT, two recent ViT attacks designed for transfer. In our white-box benchmark, rollout and GAE agree in zero of six FPR cells and three of five estimable ATT cells. The probe averages attack success rate (ASR) over four models, with 200 images per cell. SSIM is relative to the clean image. ATT is near its plateau at 4/255, while FPR gains 6.6 points at 8/255.

AttackASR at 4/255ASR at 8/255Δ (points)Median SSIM at 4/255
PGD0.9510.970+1.90.968
MIM0.9420.955+1.20.958
SAGA0.9200.950+3.00.986
TGR0.6230.772+15.00.941
FPR0.9190.985+6.60.954
ATT0.9260.958+3.10.950

Changes in the revision. Section 5, Tables 3 and 5, and Appendices A, B, G, and Q cover FPR and ATT. Appendix M qualifies the common budget.

We hope these responses address the reviewer's concerns. If so, we would be grateful if the reviewer would consider updating the score. We are happy to provide further clarification.

Good

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

In this paper, the authors proposed ViTViz, an open-source tool for exploring attention under adversarial attacks. ViTViz can enable researchers to independently investigate key dimensions of attention on any plain ViT classifier with a single class token, in the timm or Hugging Face layout. Moreover, the authors propose a key observation that different attention aggregation can lead to substantially different conclusions about adversarial attention redistribution.

Critical strengths
  1. The proposed visualization tool ViTViz is useful for the research community to analyze the adversarial attacks on the ViT.
  2. Plenty of interesting observations are provided on how attention redistribution affects the analysis of adversarial attacks on the ViT.
Critical weaknesses
  1. The core observation that "different attention aggregation can lead to substantially different conclusions about adversarial attention redistribution" is questionable.
  2. Lack of technical contribution. While some observations are interesting, the authors do not provide in-depth analysis.
Questions for authors
  1. The core observation of this paper is "Different attention aggregation operators can lead to substantially different conclusions about adversarial attention redistribution." However, the selected aggregation methods differ essentially. Rollout relies solely on the attention matrix, whereas GAE utilizes the attention matrix, gradients, and a specified class. Therefore, they are not merely different ways of aggregating the same information. GAE introduces entirely different, additional information. Thus, here is a question: Does the significant difference observed stem from attention aggregation, or from the additional gradient and class information utilized by GAE? Moreover, Rollout examines the propagation of raw attention, whereas GAE examines class-specific, gradient-based attribution. This implies that the authors might actually be comparing two distinct objects of explanation, rather than two different measurement tools for the same underlying attention redistribution.

  2. The authors uncovered plenty of interesting phenomena. For example, when using ViT-S/16, the ratio for GAE is less than 1, whereas for Base/Large models, it is frequently greater than 1. Furthermore, on the GTSRB dataset, GAE and Rollout even yield contradictory results. These findings could have served as the starting point for valuable research questions: How does model capacity alter the relationship between Gradient-based Attention Redistribution (GAE) and attack success? How might this shift be linked to factors such as attention entropy, gradient sensitivity, attention saturation, or classification margin? However, the authors merely report these phenomena without providing further explanation. If the primary contribution of an ICLR paper is to understand model behavior, one would expect it to go beyond simply reporting phenomena and to propose and validate the underlying mechanisms that drive them. In its current form, the paper leans more toward an empirical observation than a comprehensive mechanistic analysis. Therefore, the technical contribution of this paper is quite limited.

  3. The experiments compare only two interpretation methods: Rollout and GAE. However, the existing literature on ViT interpretability includes more approaches, such as Gradient-based Attention Rollout, LRP, Attention Flow, and LeGrad. Given that the authors' central argument concerns the dependence of attention interpretation on the aggregation operator, comparing only two highly distinct methods makes it difficult to determine whether this phenomenon is a general one.

1 comment

We thank Reviewer KrdV for the thoughtful review and for finding ViTViz useful to the research community and our observations interesting.

W1 and Q1 (What drives the disagreement). Rollout follows how attention mixes tokens. GAE weights that mixing by gradients from the class score. These readings can answer different questions about the same image. We varied design choices on the same attacked images. On raw attention, GAE’s aggregation gives rollout up to a constant. Across-layer gradient weighting accounts for 6.5 of the 12-cell gap. For GAE versus LeGrad, using each layer’s logit raises agreement from 16 to 22 of 24 cells and removes all four opposite-direction verdicts. For rollout versus GMAR, the residual term and product order account for 5.2 and 2.5 cells of the eight-cell gap. Each pair has a different driver.

Changes in the revision. Appendix P gives the three decompositions. Section 5 introduces them.

W2 and Q2 (Depth, capacity, and mechanism). We measured the contrast by depth and patched late attention. With common max normalization, the gradient-weighted attention of the last three layers separates 20 of 24 cells, compared with 19 for full GAE. Restoring clean late-layer patterns recovers 12%–39% of the lost margin and undoes 2%–14% of flips. Restoring the full output recovers 75%–94%. Inserting a pattern from another clean image lowers the true-class probability at least as much as the attacked pattern in five of six cells. The evidence supports a partial role for late attention. The mechanism of the flip remains open.

In four ViT-S cells from the original five attacks, GAE separates below 1 and LeGrad above 1 with the same class anchor. Using each layer’s logit removes all four reversals. The apparent capacity pattern depends on how we read attention. The PGD reversal on GTSRB also failed to recur in the three additional checkpoints with estimable results.

In a separate exploratory check for this response, we stratified by per-cell quintiles for the four factors you suggested. GAE still separates 17 of its 20 cells for classification margin, 18 for attention entropy, 19 for saturation, and 16 for gradient sensitivity. The median retained log-effect is 80%–98%. The bins can leave residual differences in each factor.

Changes in the revision. Section 6.2 and Appendix P qualify the capacity result. Appendices Q and R add the depth and patching tests. Section 6.3 and Appendix K report the checkpoint comparison.

Q3 (Additional methods and implications). We added LeGrad and gradient-based attention rollout in the form of GMAR. Each differs from GAE in 8 of the 24 estimable cells for the original attacks, although all three use the same class anchor. With a random anchor, LeGrad retains 6 of 22 separations in the same direction and GMAR retains all 15 of its original separations. GMAR’s per-image displacement changes by a median of 1.1%. Across the four operators, no pair agrees on more than 16 of 24 verdicts. We recommend reporting several complementary readings to distinguish robust findings from effects specific to an operator. The benchmark gives each model, attack, operator, and outcome its own reference. ViTViz supports this check case by case for rollout and GAE (Section 4).

Changes in the revision. Tables 3 and 11 add LeGrad and GMAR. Appendix J reports the controls and Appendix P explains their differences.

We hope these responses address the reviewer's concerns. If so, we would be grateful if the reviewer would consider updating the score. We are happy to provide further clarification.

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

This paper proposes ViTViz, an interactive framework for visualizing and statistically analyzing attention redistribution in Vision Transformers under adversarial attacks. ViTViz combines two attribution methods, Attention Rollout and Generic Attention-model Explainability, with three metrics, Wasserstein-1 distance, Jensen–Shannon divergence, and absolute entropy change, to quantify differences between clean and adversarial attention. The authors further build operator-specific reference distributions to contextualize attention displacement via percentile rankings, compare successful and failed attacks, and use variance decomposition to assess the contributions of image identity, attack choice, and attribution operator. Experiments on ImageNet-1K images with six ViT configurations and five attacks, plus a case study on GTSRB, reveal substantial operator-dependent differences in measured displacement, and show that image identity explains more of the variation than attack choice.

Critical strengths
  1. ViTViz integrates five adversarial attacks, two attribution operators, and multiple displacement metrics into a unified interactive framework. It supports side-by-side comparison of clean and adversarial attention maps and contextualizes individual measurements through benchmark-based percentile rankings.

  2. The displacement metrics capture complementary aspects of attention change: Wasserstein-1 distance reflects spatial transport, Jensen–Shannon divergence measures distributional difference, and absolute entropy change characterizes shifts in attention concentration. Controlled synthetic examples in Appendix D further illustrate their distinct sensitivities, making the metric choice well motivated.

  3. The experiments cover six ViT configurations and five attacks, and use bootstrap confidence intervals and variance decomposition to analyze operator-dependent effects. The paper also examines the influence of class anchors, normalization choices, and sample-size thresholds, and includes a GTSRB case study to assess whether the findings extend beyond ImageNet.

Critical weaknesses
  1. The practical benefits of ViTViz over existing analysis tools remain insufficiently evaluated. Although the Introduction section identifies differences from prior tools, the paper provides limited comparative evidence of how these capabilities improve the effectiveness or efficiency of adversarial attention analysis.

  2. The generalizability of the reported conclusions about operator dependence remains unclear, as the empirical evaluation is limited to Attention Rollout and GAE.

  3. The evaluation is limited to five adversarial attack methods, with insufficient coverage of recent techniques. This limits the evidence for the generalizability of the reported findings across different attack methods.

  4. The interpretability of benchmark percentiles under modified attack configurations is unclear. Although the authors provide mismatch warnings and examine CPU–GPU consistency in Appendix L (Table 9), the reference distributions are fixed, and the validity of cross-configuration comparisons is not evaluated.

Questions for authors
  1. What empirical evidence supports the practical benefits of ViTViz compared with existing visual analytics systems such as AttentionViz [1] and AdvEx [2], particularly in terms of analytical effectiveness and usability?

  2. When Rollout and GAE yield conflicting conclusions about attention redistribution, how does ViTViz help researchers interpret these differences beyond reporting the corresponding displacement metrics and percentile rankings?

  3. Table 2 shows that Rollout and GAE agree in statistical significance decisions in only 6 of the 24 estimable cells. Given that only these two operators are evaluated, what evidence supports the applicability of the reported findings to other attention attribution methods?

  4. The evaluation includes five untargeted white-box attacks (FGSM, PGD, MIM, TGR, SAGA). Would the reported operator-dependent findings also hold under targeted attacks or different attack optimization strategies?

  5. Table 6 in Appendix J shows that the proportion of significant GAE comparisons drops by about 0.2 after stratifying by the final ground-truth class probability. Could the authors quantify how much of this drop is due to changes in effect size rather than smaller sample sizes?

  6. When users modify attack configurations, how should benchmark percentiles computed from reference distributions under fixed settings be interpreted, and how can potentially misleading comparisons be avoided?

  7. Table 3 in Section 6.3 reports opposite trends in the flipped-to-unflipped displacement ratio under PGD on GTSRB for Rollout (1.109) and GAE (0.840). Since this result comes from a single model checkpoint, does the reversal also occur across other checkpoints or experimental settings?

[1] AttentionViz: A Global View of Transformer Attention. IEEE Transactions on Visualization and Computer Graphics 2023.

[2] Panda or not panda? Understanding adversarial attacks with interactive visualization. ACM Transactions on Interactive Intelligent Systems 2025.

2 comments

We thank Reviewer GoLN for the thoughtful review and for highlighting the three complementary measures and the experimental controls.

W1 and Q1 (Tool comparison). We compared five tools. Bluff and AEVis inspect attacked CNNs. ViTViz follows ViT attention during an attack and benchmarks its displacement. Usability and analytical efficiency remain untested.

Question an analyst can askAttentionVizAdvExBluffAEVisViTViz
What do a transformer's heads do across a clean corpus?YesNoNoNoNo
What does each head of a ViT attend to, layer by layer, in one clean image?YesNoNoNoYes
How does an evasion attack change a CNN's accuracy and embeddings?NoYesNoNoNo
Which neurons and connections inside a CNN does an attack change, layer by layer?NoNoYesYesNo
How does a ViT's attention map change along the attack, step by step?NoNoNoNoYes
Is that change large for its model, attack, operator, and outcome?NoNoNoNoYes
Does the answer depend on how attention is read?NoNoNoNoYes

Changes in the revision. Section 2 adds Table 1. Appendix M states the evaluation limits.

W2 and Q3 (Operator dependence). We added LeGrad and GMAR. All four operators compare mean displacement for flipped and unflipped images from the same attack runs. Of 24 estimable cells for the original attacks, rollout separates 8, GAE 20, LeGrad 24, and GMAR 16. No pair agrees on more than 16 verdicts. The operator can even reverse the direction on the same images. We recommend using several readings to show which findings carry across operators. At equal loss, LeGrad retains 11 of 24 separations and GMAR 6 of 16. At equal final probability, they retain zero and two.

Changes in the revision. Table 3 adds the operators. Appendix J tests both probability controls.

W3 and Q4 (Attack coverage). We added FPR, ATT, and targeted PGD. Rollout, GAE, LeGrad, and GMAR separate 1, 5, 6, and 3 of 6 FPR cells and 2, 4, 5, and 3 of 5 ATT cells. Rollout and GAE agree in 9 of 35 untargeted cells and 3 of 6 targeted PGD cells. Ratios compare mean displacement for target hits and misses. Bold marks rejection of a ratio of 1 at 5%. Parentheses count separations if any prediction change is a success.

ModelHit rateRolloutGAELeGradGMAR
ViT-S/1675%1.0510.8341.1810.931
ViT-S/3292%1.0410.9531.2711.014
ViT-B/1688%0.9371.1921.2420.982
ViT-B/3285%1.0521.0231.1840.965
ViT-L/1637%1.0531.1231.1821.162
ViT-L/3236%1.1061.2071.1271.131
Cells that separate (of 6)1 (1)4 (4)6 (5)3 (3)

Changes in the revision. Section 3 broadens the threat model. Section 5, Table 3, and Appendices A, B, G, and Q add FPR and ATT. Appendix O reports targeted PGD (Table 16).

W4 and Q6 (Percentile drift). We varied the budget across four operators and seven attacks. Median percentiles move from 48–54 at 4/255 to 1–25 at 1/255 and 62–94 at 16/255. Outside the reference configuration, the percentile ranks the run against the fixed benchmark. ViTViz flags the departure (Section 4).

Changes in the revision. Appendix A reports percentile drift.

Q2 (Reading a disagreement). We show how the operator switch and slider locate when and where rollout and GAE diverge on the same run. Operator-specific percentiles show how unusual the final displacement is. The raw layer and head views show the late layers where the gradient-weighted contrast is strongest.

Changes in the revision. Section 4 adds the workflow and cites Appendix Q’s depth analysis.

Q5 (Effect size and sample size). We replaced quintiles with bands of 0.1 and used size-matched random subsamples. At equal final probability, GAE retains zero of 20 separations in the original direction. Six cells lose common support, about 7.4 are lost to sample size, and about 6.6 to a smaller effect. At equal loss, GAE retains 10, with losses of 1, 5.4, and 3.6. The submitted Table 6’s drop from 20 to 16 was for equal loss.

Changes in the revision. Appendix J reports the bands (Table 10) and the subsampling analysis.

Q7 (GTSRB replication). We tested five more checkpoints. None of the three with estimable PGD results reproduced the reversal. Ratios are 1.15–1.52 for rollout and 1.23–1.30 for GAE. Two checkpoints fall below the sample floor.

Changes in the revision. Section 6.3 and Appendix K report all six checkpoints (Tables 4 and 12).

We hope these responses address the reviewer's concerns. If so, we would be grateful if the reviewer would consider updating the score. We are happy to provide further clarification.

I am grateful to the authors for their response. Although the paper provides a comprehensive evaluation of attention redistribution in vision transformers under adversarial attacks, the proposed method combines existing techniques into a unified framework, and its technical contribution beyond this integration remains unclear. Another reviewer raised the same concern. In addition, the experimental analysis offers limited support for the proposed mechanism. I therefore maintain my original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Thank you for the feedback on ViTViz, the three complementary measures, and the experimental controls. The revised manuscript marks changes in blue.

  • Operators and attacks. Table 3 adds LeGrad, GMAR, FPR, and ATT. In 35 estimable cells, rollout separates 11, GAE 29, LeGrad 35, and GMAR 22. Rollout and GAE agree in 9 of 35 cells. Under separate targeted PGD, they agree in 3 of 6 (Appendix O). In the shared rerun, nine ratios from the submitted Table 2 change by at most 0.003. The original image counts and rollout and GAE verdicts are unchanged.
  • Operator differences. The largest Shapley shares are 6.5 cells for across-layer gradient weighting (rollout to GAE), 11.0 for the choice of logit (GAE to LeGrad), and 5.2 for the residual term (rollout to GMAR). With common max normalization, the last three layers alone separate 20 of 24 cells, compared with 19 for full GAE. Restoring their clean attention patterns recovers 12%–39% of the lost margin (Appendices P–R).
  • Test and replication. With 20 unflipped images, the Welch-statistic permutation test rejects about 5% of simulated null draws, compared with 8.5% for the percentile bootstrap. The PGD reversal did not recur in the three additional GTSRB checkpoints with estimable PGD results. Rollout and GAE agree in 20 of 26 estimable cells across the five public checkpoints (Sections 5 and 6.3, Appendices K and N).
  • Tool and controls. Table 1 compares five tools. For the same reference at 4/255, the median percentiles are 48–54 at 4/255, 1–25 at 1/255, and 62–94 at 16/255 (Appendix A). Appendix J uses bands of 0.1. At equal final probability, GAE retains zero of 20 original contrasts in the same direction, with losses of 6 to common support, about 7.4 to sample size, and about 6.6 to a smaller effect.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.