acceptodds
Under review as a conference paper at ICLR 2027

Robust multi-modal models by explicitly optimizing uniformity and alignment

Abstract

We define the robustness of a multi-modal embedding using the margin of one nearest neighbor retrieval and derive the form of the optimal robust embedding (ORE) for various finite dataset sizes () and dimensions (). Specifically, we show that for the most robust embedding is one where the modalities are aligned and the embeddings form the vertex of a regular simplex. Previous work has shown that the ORE is also the global optimum of the CLIP loss yet gradient-based optimization of the CLIP loss does not converge to the ORE, and instead learns embeddings with a modality gap. To alleviate this failure, we analyze an alternative loss, Explicit Multi-Modal Uniformity and Alignment (EMMUA), whose global optimum is also the ORE. We show both theoretically and experimentally that gradient-based optimization on the EMMUA does not learn a modality gap and finds significantly more robust embeddings than those obtained with the CLIP loss.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
robustness, multi modal models, modality gap, contrastive learning, clip

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EmrbOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the relation between modal gaps and retrieval robustness in multimodal contrastive learning. The optimal robust embedding is defined by the average cosine gap of cross-modal nearest neighbor retrieval. It is proved that when the number of sample pairs satisfies N≤d+1N \le d + 1, the configuration where the two modalities are perfectly aligned and the embeddings form a regular simplex has the optimal average gap, and it is consistent with the global optimal configuration of the CLIP loss. Based on this, the paper focuses on the existing explicit uniformity and alignment loss EMMUA, analyzing its global optimal solution and gradient optimization dynamics, and attempting to explain the mechanism by which it alleviates modal gaps and improves embedding robustness. The verification process includes directly optimizing the synthesized embedding vectors, training the model on MAD, CIFAR-100, and ImageNet, fine-tuning on MS-COCO, and comparing the performance under modal gaps, retrieval gaps, and different perturbations. The results show that EMMUA can significantly reduce modal gaps, increase retrieval gaps, and improve the resistance of embeddings to Gaussian noise and quantization errors, but the clean performance on some datasets has decreased, and the improvement for input image perturbations is relatively limited.

Critical strengths
  1. This paper focuses on the modal gap in multimodal models, paying attention to its impact on retrieval stability and embedding quantization. It links the geometric properties of the representation space with actual retrieval requirements.
  2. The cosine gap between correctly paired samples and the hardest negative samples is used to measure retrieval stability. Under the condition of N≤d+1N \le d + 1, the optimal average gap is linked to the complete alignment rule simplex, providing a clear theoretical reference for evaluating embedding quality.
  3. This paper connects the optimal configuration of the loss function, the gradient training dynamics, and the final retrieval robustness, which helps to understand why the objective function has an ideal optimal solution, while the actual training may still retain the modal gap.
Critical weaknesses
  1. Theorem 7 uses the EMMUA loss without the denominator positive sample term. However, Appendix A.4 introduces another form that includes this term and writes both in Eq. 51 as being equal. The difference between the two losses is not constant, and the gradients may also be different. Therefore, the current dynamics proof cannot directly cover the actual training objective.
  2. The proof of modal gap contraction does not adequately handle the spherical constraint. Theorems 3.6 and the Appendix are mainly derived based on vector updates in the Euclidean space, while the theoretical issues and some experiments limit the embedding to the unit sphere. Normalization or projection onto the tangent space will change the update direction, and the relevant conclusions need to be re-proven. Additionally, the reduction of modal mean distance does not equal the complete alignment of paired samples, and it cannot directly lead to global convergence.
  3. Theorem A.9 gives an upper bound for the embedding norm, but its proof actually requires a positive lower bound, so the current input robustness guarantee does not hold. Theorem 3.5 claims that the conclusion applies to any λ\lambda, but it also omits the necessary condition: when λ=0\lambda=0, the two modalities do not have the constraint of having to be aligned with each other.
  4. The theoretical definition considers the average margin of all queries, but Fig. 5 and Table 2 mainly count the samples that are correctly predicted by each model. Since the sets of correct samples for the two models are different, this comparison is not sufficient to directly prove that the average margin and robustness on the complete test set have achieved the same improvement.
Questions for authors

See Critical Weaknesses.

2 comments

We thank the reviewer for their thoughtful comments.

  1. Loss Consistency - Thank you for pointing this out. The objective we train and analyze is Eq. 7, which has no positive-pair terms in the denominators. The inconsistency was in the restatement of the loss in Appendix A.4. We replaced it with Eq. 7 itself, so that all of the dynamics analysis is derived for the trained objective. The gradients in Eq. 8 and Eq. 50 were already those of Eq. 7, so Theorem 3.6 is unchanged.

  2. Spherical Constraint - we agree that our analysis is of a simplified setting of Euclidian space without a spherical constraint. However, as [2] showed in high dimensions the spherical and Euclidian analysis are rather similar. We also follow previous works [2,3,4] that perform similar analysis without the spherical constraint. Finally, we provide empirical evidence that the same phenomena seen in euclidian space (figure 2) occurs on the sphere (all other figures).

  3. We agree that a vanishing mean gap implies neither pairwise alignment nor global convergence, and we do not claim global convergence and will state this explicitly. The implication runs the other way: every perfectly aligned solution has zero gap. We therefore also report pairwise alignment directly:

    DatasetLossAlignment (mean true-pair cosine sim.)
    MADCLIP0.736
    EMMUA0.958
    CIFAR-100CLIP0.568
    EMMUA0.609
    ImageNetCLIP0.349
    EMMUA0.695
    MS-COCO (finetune)CLIP0.297
    EMMUA0.520

    Alignment is higher with EMMUA on every dataset. The gain on CIFAR-100 is modest, although the gap there falls from 0.356 to 0.043. As discussed in Section 4.2, with one caption per class the optimum maps all images of a class to one point, and alignment is limited by how far the image encoder collapses each class. These numbers will be added to Table 1.

  4. Theory:

    1. Theorem A.9 - Thank you for the correction. We corrected this in the manuscript. As you note, this concerns only the theoretical extension, not the embedding noise results.

    2. Theorem 3.5 - Thank you for pointing this out. This is supposed to be for any λ>0\lambda > 0 as both you and reviewer v2WA pointed out. We fixed this in the manuscript.

  5. Average margin vs Correctly classified margin - this is indeed a delicate point. To compare the two models on identical sets, we additionally report a margin defined for every query: with j=arg⁡max⁡k⟨xi,yk⟩j = \arg\max_k \langle x_i, y_k \rangle, the top-1 minus top-2 similarity ⟨xi,yj⟩−max⁡k≠j⟨xi,yk⟩\langle x_i, y_j \rangle - \max_{k \neq j} \langle x_i, y_k \rangle. It coincides with the margin definition on correctly retrieved queries and measures how stable the retrieved result is on the others.

    DatasetCLIP mean top-1 − top-2EMMUA mean top-1 − top-2
    CIFAR-1000.21630.5046
    MAD0.23230.6370
    ImageNet0.06070.2989
    MS-COCO ft0.02220.0657

    EMMUA's value is ~2-5× larger on every evaluation set, including the three on which the MS-COCO-finetuned models are evaluated zero-shot. We present this as a measure of retrieval stability over the full set, not of correctness: it is non-negative by construction and is also larger for EMMUA on incorrectly retrieved queries. It is consistent with Table 5, where retrieval consistency under noise is higher for EMMUA on every dataset.

[1] "Decoupled contrastive learning", Yeh et al., 2022

[2] "Is the Modality Gap a Bug or a Feature? A Robustness Perspective", Chowers et al., 2026

[3] "Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data", Zhang et al., 2024

[4] "Explaining and Mitigating the Modality Gap in Contrastive Multimodal Learning", Yaras and Chen et al., 2026

The authors answered my questions, and I maintain my original acceptance.

Reviewer_z3UYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors investigate the multimodal robustness through nearest-neighbor retrieval margins. This work connects the optimal robust embedding to a regular simplex and analyzes Explicit Multi-Modal Uniformity and Alignment as an alternative to the CLIP loss. The experiments examine the modality gap and robustness to embedding noise and quantization.

Critical strengths
  • This work connects retrieval margins with embedding geometry and optimization dynamics. This gives an informative explanation of why a loss's optimum and its learned representations can differ.

  • The loss comparisons control initialization and training settings. Table 2 reports substantial gains under embedding perturbations; for example, with two-level quantization on MS-COCO, accuracy is 37.8% for Explicit Multi-Modal Uniformity and Alignment and24.5% for the CLIP loss.

Critical weaknesses
  • The novelty is incremental. The loss itself comes from Yin et al. (2026b), the simplex optimum of the CLIP loss and the gap dynamics build on Chowers et al. (2026) and earlier work, so the contribution is an analysis of a known loss.

  • Clean accuracy drops on CIFAR-100 (59.3 to 56.3) and MS-COCO (43.0 to 40.9), and the discussion admits that lambda and tau need study before training at scale. The claim of "significantly more robust embeddings" is therefore shown only in a regime far from the one that motivates the paper.

  • The dynamics analysis needs a consistent objective. Appendix A.4 introduces positive-pair terms in the denominators, but Equation 51 equates that expression with one omitting those terms. The expressions are not algebraically identical and have different gradients. A derivation for the actual objective in Equation 7 is needed to support Theorem 3.6.

  • Theorem A.9 assumes that embedding norms are bounded above, while the step from Equation 49 to Equation 50 requires a positive lower bound on the norms. The input robustness statement is therefore unsupported as written. This affects the theoretical extension, rather than invalidating the reported embedding-noise results.

Questions for authors

Please address the points raised in the critical weaknesses section.

2 comments

We thank the reviewer for their helpful comments.

  1. Novelty - Our paper isn't the first to provide theoretical analysis of existing losses, such as [1,2,3]. And while our theoretical results certainly build on previous work, we believe our contributions are more than incremental. We are the first to show that the optimal robust embedding is the same as the embedding that minimizes the CLIP loss and the EMMUA loss, and we are the first to show analytically that the dynamics of the EMMUA loss are superior to the dynamics of the CLIP loss under identical assumptions.

  2. While table 1 does show that for some datasets clean accuracy is lower with EMMUA than it is with the CLIP loss, this is by no means an intrinsic property of the two losses but rather seems to depend on the architectures and hyperparameters (e.g. for ImageNet, clean accuracy is lower with CLIP). We disagree with your claim that "The claim of 'significantly more robust embeddings' is therefore shown only in a regime far from the one that motivates the paper". Rather our paper shows that EMMUA embeddings consistently have much higher margins than CLIP embeddings, across a range of regimes (figures 1,3,4,5 and tables 2,4,5).

  3. Loss Consistency - Thank you for pointing this out. The objective we train and analyze is Eq. 7, which has no positive-pair terms in the denominators. The inconsistency was in the restatement of the loss in Appendix A.4. We replaced it with Eq. 7 itself, so that all of the dynamics analysis is derived for the trained objective. The gradients in Eq. 8 and Eq. 50 were already those of Eq. 7, so Theorem 3.6 is unchanged.

  4. Theorem A.9 - Thank you for the correction. We corrected this in the manuscript. As you note, this concerns only the theoretical extension, not the embedding noise results.

[1] "Decoupled contrastive learning", Yeh et al., 2022

[2] "Is the Modality Gap a Bug or a Feature? A Robustness Perspective", Chowers et al., 2026

[3] "Explaining and Mitigating the Modality Gap in Contrastive Multimodal Learning", Yaras and Chen et al., 2026

Reviewer_z3UYOpenReview ↗

Response to the Authors

Thank you for the rebuttal. However, it did not fully address my concern about the novelty. Therefore, I maintain my original score.

Reviewer_v2WAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the relationship between multimodal embedding geometry and retrieval robustness. It formulates robustness through the average nearest-neighbor retrieval margin and shows that, for N≤d+1N \leq d+1 , the optimal robust embedding corresponds to perfectly aligned modalities arranged as a regular simplex. Building on an existing Explicit Multi-Modal Uniformity and Alignment (EMMUA) objective, the authors analyze how its gradient dynamics avoid modality-gap behavior observed with CLIP training. Experiments demonstrate larger retrieval margins and improved robustness to embedding-space Gaussian noise and quantization, with broadly comparable clean performance. However, improvements under input-space perturbations are limited.

Critical strengths
  1. The paper establishes a useful connection between nearest-neighbor retrieval margins and the geometry of multimodal embeddings. The regular-simplex characterization for N≤d+1N \leq d+1 provides a theoretically interpretable benchmark for optimal robustness and helps clarify the relationship between alignment, separation, and retrieval stability.
  2. Insightful analysis of optimization dynamics. The distinction between a loss having a desirable global optimum and gradient-based training actually reaching that optimum is important. The analysis of how EMMUA avoids certain modality-gap-preserving dynamics offers a useful perspective on the optimization limitations of conventional multimodal contrastive learning.
  3. The experiments demonstrate substantially larger margins and improved resistance to Gaussian noise and quantization compared with CLIP, across multiple datasets.
Critical weaknesses
  1. Insufficient comparison with closely related works. The EMMUA objective is adapted from Yin et al. (2026), while Chowers et al. (2026) already investigate the relationship between modality gaps and robustness. Previous work, including Fahim et al. (2024) and Yaras et al. (2025), also examines explicit uniformity/alignment and contrastive optimization dynamics. Although the retrieval-margin formulation and specific theoretical analysis may be original, the paper does not sufficiently distinguish its central contributions from these existing results. In particular, it is unclear how much of the claimed optimization advantage follows directly from properties of the previously introduced objective.
  2. Theorem 3.2 states that a retrieval margin δ\delta guarantees robustness under perturbations satisfying $ \Vert v \Vert <2\delta$ . However, the unit-norm assumptions provide the general sufficient bound$ \Vert v \Vert <\delta / 2$, making the stated robustness guarantee incorrect. Additionally, Theorem 3.5 claims uniqueness of the aligned regular-simplex solution for any λ\lambda, although this does not hold for λ=0\lambda=0, when the modalities can independently minimize their uniformity objectives. These issues need correction because they affect the precision of the paper's theoretical contributions.
  3. The experiments primarily compare EMMUA with CLIP and a post-processing modality-shift correction. However, closely related methods already address modality-gap reduction through uniformity regularization, modified negative sampling, and alternative training objectives. Comparisons with approaches such as Fahim et al. (2024), and xNCE (Mager et al., 2026) are needed to determine whether EMMUA offers distinctive advantages over existing methods.
  4. The main optimality results for CLIP, EMMUA, and ORE rely on N≤d+1N \leq d+1 . In practical multimodal training and retrieval, the number of examples typically greatly exceeds the embedding dimension. Although the authors provide empirical results for larger N, the theoretical equivalence between the objectives and optimal robustness remains unestablished in this regime. Furthermore, the formulation assumes designated one-to-one positive pairs, whereas realistic image–text datasets often contain multiple semantically valid matches.
  5. EMMUA does not consistently preserve clean performance. For example, CIFAR-100 accuracy decreases from 59.3% to 56.3%, and MS-COCO R@1 decreases from 43.0% to 40.9%. The paper lacks a systematic analysis of how the alignment weight λ\lambda and temperature τ\tau influence these trade-offs. Comparisons at matched clean accuracy would make it easier to determine whether the robustness improvements reflect a genuinely superior operating point.
  6. The evaluation does not sufficiently establish whether these geometric changes improve broader multimodal representation quality. Prior work has also questioned whether eliminating the modality gap is universally beneficial. Evaluations involving compositional retrieval, hard negatives, and downstream transfer would help demonstrate benefits beyond the specific geometric and numerical perturbation metrics optimized by the method.

References:

  • Mager, Fabian, Hiba Nassar, and Lars Kai Hansen. "On the modality gap and the contrastive loss in multi-modal representation learning." arXiv 2026
  • Yaras, Can, Siyi Chen, Peng Wang, and Qing Qu. "Explaining and mitigating the modality gap in contrastive multimodal learning." Conference on Parsimony and Learning 2025.
Questions for authors
  1. Can the authors revisit Theorem 3.2? on the claimed perturbation bound. Additionally, does the uniqueness result in Theorem 3.5 require λ>0\lambda>0?
  2. How does EMMUA compare against existing training-time alternatives, including explicit uniformity regularization and xNCE, when controlling for architecture, training budget, and clean retrieval performance?
  3. What evidence supports the connection between EMMUA and optimal retrieval robustness when N≫dN\gg d ?
  4. Given the relatively modest improvements under input-space corruptions, what practical scenarios motivate the proposed objective beyond embedding quantization and numerical perturbations? and broader robustness claims?
2 comments

Thank you for the helpful comments. We agree in general that there is much more to be said about the loss - comparisons with Fahim et al, Mager et al and Yaras et al, as well as more experiments on downstream settings. However, our paper tries to focus on providing the theoretical justification for the use of this loss in the first place, and its relation to embedding space robustness in the context of retrieval. We do agree that the paper would benefit from wider comparisons.

To answer your specific questions:

  1. Theoretical Comparison to related works - We are not the first to train multi-modal models with explicit unimodal uniformity: Fahim et al. use it as a regularizer added to the CLIP loss, and Yin et al. propose it as a loss of its own, primarily for more than two modalities. What those papers do not provide, and what we contribute, is the following:

    1. A robustness criterion and its solution. We define the optimal robust embedding through the average retrieval margin and prove that for N≤d+1N \leq d+1 it is the aligned regular simplex (Theorem 3.3). This is a statement about the max-margin problem itself; it does not follow from the known optimum of the CLIP loss.

    2. Dynamics. Chowers et al. [1] show that the gap harms robustness and analyze why the CLIP loss preserves it. We show that EMMUA shrinks the gap under each of those same conditions (Theorem 3.6). To the reviewer's question of how much follows directly from the objective: while Chowers et al. show that CLIP's gradient creates and perserves an orthogonal gap, we show that EMMUA's gradient do no such thing, eliminating any harm to robustness caused by the existence of the gap.

    3. Evidence that the remedy must be in training. Aligned random points have larger margins than pretrained models, including adversarially trained ones (Fig. 1d), and post-hoc shifting [1] does not recover the margin (Table 2). We show that training on EMMUA achieves superior robustness and margins on various datasets, coinciding with out theory.

    We will add a discussion of Yaras et al. [2] and Mager et al. [4] to Section 2.

  2. Benchmarking - We agree, and have added two training-time baselines: LCUAL_{CUA} (Fahim et al., 2024) and xNCE (Mager et al., 2026). Both are trained with the same architecture, initialization, seed, schedule and budget as our CLIP and EMMUA runs. The hyperparameters of LCUAL_{CUA} and xNCE follow their respective papers. Both baselines are hybrids of the two losses we analyze. LCUAL_{CUA} is exactly the sum of CLIP's loss with EMMUA. xNCE adds intra-modal negatives to the CLIP denominator, so its gradient contains EMMUA's intra-modal repulsion alongside CLIP's cross-modal repulsion. Our analysis therefore predicts that both reduce the gap, but since both keep the cross-modal repulsion term that pushes yiy_i away from the other modality, their margins should be smaller than EMMUA's, which removes that term. This is what we observe.

    Directly optimized embeddings (setting of Fig. 1, N = 50, d = 3).

    LossMean true-pair cosine sim.Mean marginFrac. margin > 0 (= R@1)Uniformity ↓
    CLIP−0.01670.00150.88−1.2096
    EMMUA1.00000.10001.00−2.2058
    xNCE (rmix=0.5r_{mix}=0.5)1.00000.08381.00−2.1715
    LCUAL_{CUA} (Fahim et al.)0.99180.00450.66−1.6616

    EMMUA and xNCE both reach perfect alignment, with EMMUA attaining the largest margin. LCUAL_{CUA} reaches high pairwise cosine similarity (0.99), but its margins remain close to CLIP's.

    Finetuning on MS-COCO (same protocol as Table 2).

    LossClean R@1Quant L=2Quant L=4Quant L=16Gauss σ=0.1Gauss σ=0.15Gauss σ=0.2Margin (correct)
    CLIP43.024.524.526.17.22.21.00.036
    EMMUA40.937.837.740.730.819.810.50.102
    xNCE (rmix=0.5r_{mix}=0.5)40.432.332.438.819.59.24.20.044
    LCUAL_{CUA}38.131.731.735.023.514.17.70.054

    All three alternatives are more robust than CLIP, confirming that intra-modal uniformity is the relevant ingredient. EMMUA is the most robust at every noise level: it has roughly twice the margin of the next-best loss (0.102 vs 0.054) and retains 92% of its clean R@1 under 1-bit quantization, against 80% for xNCE, 83% for LCUAL_{CUA} and 57% for CLIP. The comparison with xNCE is at essentially matched clean accuracy (40.9 vs 40.4 R@1). Across the four losses, accuracy under Gaussian noise is ordered exactly as the mean margin, in line with the margin-based view of robustness in Section 3. We will add these results and expand them to all datasets for the final manuscript.

  1. Theoretical claims:

    1. Theorem 3.5 - Thank you for pointing this out, you are correct and this indeed should be when λ>0\lambda > 0, we've corrected this in the manuscript.

    2. Theorem 3.2 - We believe you might have misread - the theorem currently explicitly states that ∥v∥≤δ/2\lVert v \rVert \leq \delta/2, not 2δ2\delta.

  2. Restrictive theoretical regime - First, note that we do not assume that the pairs are one-to-one (we do assume that there exist x_i and y_i but there can be y_i=y_j which is the case for almost all our experimental datasets such as CIFAR10 and ImageNet). Regarding N>d+1N>d+1 , in most practical applications the CLIP loss is optimized with mini-batches and (as shown in [4]) it can be thought of as the sum of the losses of all possible minibatches. The theory then implies that the embedding will be optimal in terms of the minibatch loss if for almost all mini-batches xi=yix_i=y_i and xix_i are close to the vertices of a regular simplex. This was already shown in the paper on the CLIP loss [4], and holds equally well for EMMUA. Additionally, our theoretical results also give a rigorous upper bound on the margin for N>d+1N>d+1 and the empirical results suggest that EMMUA optimization is close to this upper bound (figures 3,5).

  3. Input Noise - As we state in the limitations, a very interesting work would be to combine embedding space robustness with a lipschitz model, thus ensuring input robustness as well. Additionally, this might also prove useful when training with large datasets and\or augmentations - the more common approaches to ensuring input robustness.

  4. Practical scenarios - The setting our definition targets is perturbation of stored embeddings, which is routine in large-scale retrieval:

    • Quantization of vector indexes. Low-bit and binary quantization are standard for large embedding databases. Our L = 2 setting is one bit per coordinate, a 32× reduction from float32. At that rate EMMUA loses 3.1 points on MS-COCO and 0.3 on ImageNet; CLIP loses 18.5 and 8.0.
    • Approximate nearest-neighbor search and other lossy operations on embeddings, which perturb similarities directly and are absorbed by a larger margin.

[1] "Decoupled contrastive learning", Yeh et al., 2022

[2] "Is the Modality Gap a Bug or a Feature? A Robustness Perspective", Chowers et al., 2026

[3] "Explaining and Mitigating the Modality Gap in Contrastive Multimodal Learning", Yaras and Chen et al., 2026

[4] "Mini-Batch Optimization of Contrastive Loss", Sreenivasan et al., 2023

Comments on the paper

1 comment
Area_Chair_pUEPOpenReview ↗

[Open Discussion] Thank you for your review and please check authors' rebuttal

Dear All Reviewers,

We have received mixed opinions on this paper. Thank you all for your thoughtful reviews and constructive feedback.

Could you please review the authors' rebuttal and share your updated opinions on whether your concerns have been adequately addressed? Please also let us know if you have any remaining questions or concerns that require further clarification.

Dear Reviewer Emrb,

You provided a positive assessment of this paper. Could you please confirm whether the authors' rebuttal has adequately addressed your remaining concerns, particularly regarding the theoretical validity of the optimization dynamics under spherical constraints and the evaluation of retrieval margins?

Dear Reviewers v2WA and z3UY,

You both expressed reservations about this paper. Could you please check whether the authors have provided sufficient evidence and clarification to address your concerns, such as the novelty and correctness of the theoretical contributions, comparisons with relevant baselines, and trade-offs between clean performance and robustness?

Thank you all again for your time and valuable contributions to the review process.

Best regards,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.