acceptodds
Under review as a conference paper at ICLR 2027

Scaling Online Optimal Transport Minibatches for Flow Matching

Abstract

A central challenge in scaling flow models is maintaining stability in both training and integration. In flow matching, this stability depends critically on how source samples are paired with data. Independent random pairing introduces substantial stochasticity, motivating minibatch optimal transport (OT) to construct online source-target couplings that reduce this randomness (Tong et al., 2024). However, as applications extend to larger and higher-dimensional datasets, small minibatches can limit these benefits, while scaling the coupling to substantially larger minibatches at every iteration is computationally prohibitive. Moreover, how OT coupling affects the full stochastic gradient and integration remains theoretically unresolved. In this work, we develop a scalable online entropic OT approximation that enables substantially larger source-target couplings while preserving the benefits of OT. We close the theoretical gap for both unregularized and entropic OT by proving strict reductions in full stochastic-gradient variance and integration error relative to independent pairing, and further characterize how these improvements strengthen as the online coupling minibatch grows. Experiments on single-cell and image-flow training validate these predictions, showing that larger online couplings consistently improve training stability and final generation performance.

Primary area:
generative models
Keywords:
Flow models, continuous normalizing flows, stochastic gradient variance

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

This paper integrates the existing FlashSinkhorn solver into the flow matching training pipeline, making larger-scale online OT pairing computationally feasible. It theoretically analyzes how the coupling and candidate-set size affect gradient variance and integration error. The authors compare standard OT-CFM with EOT-CFM using larger pairing candidate sets on synthetic data, single-cell data, and image-generation tasks.

Critical strengths
  1. The paper provides an implementation that can be integrated into existing PyTorch training pipelines. By incorporating FlashSinkhorn, it makes larger-scale online OT pairing feasible with modest additional overhead and improves OT-CFM training and generation performance on the evaluated tasks.

  2. The paper further characterizes the theoretical relationship between OT pairing, velocity-target uncertainty, and stochastic-gradient variance, helping clarify how different pairing strategies affect gradient variance under ideal fitting conditions.

Critical weaknesses
  1. The methodological novelty is limited. The pairing objective, entropic regularization, and FlashSinkhorn solver used in this paper all come from prior work. The main methodological addition is their integration into an online training pipeline. The distinctive value of this integration relative to existing pairing methods needs to be established more clearly.

  2. The additional theoretical insights and their practical guidance are limited. The paper further explains properties of existing OT pairing methods, but its main conclusions rely on ideal fitting conditions. It does not sufficiently explain how these results can guide finite training processes, approximate coupling computation, or the choice of candidate-set size and regularization strength. Their implications for practical training and subsequent method development therefore remain unclear.

  3. The experimental coverage and comparisons are insufficient. The image-generation experiments cover only CIFAR and ImageNet-64, without validation at higher resolutions or with larger modern generative models, limiting the evidence for broader scalability. Although the paper reports some computational overhead, it lacks comparisons with other relevant pairing methods under matched conditions in terms of both quality and total training cost. Its relative advantages therefore remain insufficiently established.

Questions for authors
  1. What new insights does the theoretical analysis provide beyond existing understanding? Do these results change our understanding of the applicability or relative merits of different pairing methods, or guide the choice of pairing method, candidate-set size, and regularization strength in practical training? In particular, which conclusions remain applicable before ideal fitting is achieved?

  2. Could the authors compare the training computational costs and final generation quality of the proposed approach with other relevant pairing methods under matched conditions? Where applicable, the comparison should include the total costs of precomputation, pairing, and model training. Quality under the same total computational budget, or the computational cost required to reach the same quality, would help clarify the practical advantages of the approach.

  3. Are there results on larger models, higher resolutions, or more complex datasets that support the paper’s scalability claims? If such results are unavailable, could the authors clarify the scope of the conclusions supported by the current experiments and the potential limitations when extending the approach to these settings?

6 comments
AuthorsOpenReview ↗

Author response (1/2)

We sincerely thank you for providing these detailed and insightful points. We address each of them below:


W1, W2 & Q1: theory's practical guidance to method

  • First, we answer Q1 regarding "What new insights does the theoretical analysis provide beyond existing understanding?".

We believe the key significance lies in how the relationship between minibatch size, gradient variance, and generation error is mathematically established. Taking gradient variance as an example, as explained in line 280–283, our key contribution is to connect minibatch size to velocity uncertainty through the transport-cost residual, and then connect this uncertainty to gradient variance.

Zoom into Theorem 4.15, the gradient variance is bounded by RR (velocity uncertainty) and essential supremum JJ-norm (the largest squared model-Jacobian norm). Specifically, Jacobian norm is very large at the beginning and decreases as training progresses (see Figure 6 in Appendix B). By Propositions 4.12 and 4.14, a larger coupling minibatch can better control velocity uncertainty. This also empirically aligns with Figure 2, where the variance reduction gain is most significant during the first quarter of training. Together, these findings motivate using a large minibatch initially to suppress the uncertainty amplified by a large Jacobian, then switching to smaller minibatches as the Jacobian norm decreases.

  • Then, we move to W2 and Q1 regarding "It does not sufficiently explain how these results can guide finite training processes" and "Do these results change our understanding of the applicability...practical training?"

We address this through an adaptive scheduling strategy (EOT-CFM-scheduled). Taking an example from CIFAR, it starts with a coupling minibatch of 8192 to first and gradually reduces it to 2048. On CIFAR-100, it achieves 7.29 FID, comparable to 7.25 with EOT-CFM-8192, while requiring only 12.5 hours of training, close to the 12.2 hours of OT-CFM with 7.89 FID. On ImageNet, OT-CFM takes 55.2 h with 13.26 FID, EOT-CFM-scheduled takes 56 h with 12.96 FID, and EOT-CFM-10K takes 58.8 h with 12.94 FID. Further training-time comparisons are provided in our response to Q2.

We also add text-to-image generation experiments to demonstrate applicability to conditional generation (further details in our response to W3 and Q3). On text-conditioned CUB, EOT-CFM-scheduled achieves 9.98 FID in 18.5 hours, compared with 9.94 FID in 20.2 hours for EOT-CFM-8192 and 10.27 FID in 18.1 hours for OT-CFM.

We believe these results demonstrate a concrete benefit of scalable online coupling, which provides the flexibility to adapt the coupling context in ways that previous coupling methods could not achieve, addressing W1 and Q1 regarding your concern on "the distinctive value of this integration relative to existing pairing methods".

  • Finally, we answer Q1 for whether this "remain applicable before ideal fitting is achieved?".

We first clarify that optimality is a common setup in FM analysis, and we have expanded the discussion of prior work in lines 203–208. Optimality removes the variance contribution from imperfect fitting of the conditional-mean velocity, providing a common basis for comparing gradient variance across different couplings.

Without ideal fitting, the corresponding bound in Theorem 4.15 only includes an additional fixed-round fitting error term, and this uncertainty-amplification mechanism still remains applicable. We provide the derivation and discussion in Appendix B.

AuthorsOpenReview ↗

Author response (2/2)

W3 & Q3: experimental coverage is insufficient

We add text-conditioned flow training experiments to evaluate our method on a more complex task relevant to modern generative AI applications. We trained on CUB task, providing both quantitative FID results below and qualitative examples in Figure 4.

NFE 25NFE 100Adaptive
OT-CFM-12811.0410.6110.27
EOT-CFM-409610.8810.4410.12
EOT-CFM-819210.4910.209.94

Beyond generation quality, we find that larger minibatch context also improves text alignment. For example, in the third panel of Figure 4, the prompt is "bird with a long beak, blue eyes, and dark gray and black feathers". However, both baselines incorrectly assign blue to the feathers, while ours with a larger coupling minibatch follows the proper feather color, which shows that larger coupling minibatch may provide a more consistent conditional training signal that helps associate textual attributes with their intended visual components.


Q2: matched condition's training cost and quality

We apologize for missing this point. On CIFAR-100, OT-CFM takes 12.2 h, our largest EOT-CFM-8192 takes 14.3 h, and EOT-CFM-scheduled takes 12.5 h. Shown in Figure 5, we also train those methods for additional steps to extend their training time to 14 h, and our larger minibatch one still performs better under both same training time and same training steps. We note that the most of the additional overhead comes from coupling computation, so with larger models and more complex text-conditioned task, this overhead will account for a smaller fraction of total training time as flow optimization processes takes most of the training time. We have added these to the updated manuscript.


Again, many thanks for your careful reading and thoughtful questions. We believe your valuable feedback has helped us further strengthen the significance of our contributions, particularly by better connecting our theoretical results to their practical implications. We hope that our responses have addressed your concerns, and we look forward to an open-minded discussion if any such concerns remain.

Sincerely,

Authors of Paper 187

Hi reviewer dJhy,

As the rebuttal period comes to the end, thanks for your time to check our response to see if there's any further questions or concerns.

Reviewer_Q4VnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes choosing noise–data pairs from larger batches improves OT-based flow matching. It uses the existing FlashSinkhorn solver to recompute these pairings at every training step, while keeping the number of pairs used to update the model fixed. It also provides theoretical results connecting better pairings to lower gradient variance and lower one-step Euler integration error under idealized assumptions. Experiments on synthetic data, single-cell data, CIFAR-100, and ImageNet-64 show that larger pairing batches generally reduce gradient variance and improve generation quality, with modest reported memory overhead.

Critical strengths

The paper shows a practical way to scale online OT coupling without increasing the regression batch size. The memory overhead is small. Results on synthetic, single-cell, and image-generation tasks give useful evidence that larger online coupling contexts can help flow matching. This makes the method potentially useful for existing OT-CFM training pipelines.

Critical weaknesses
  1. Not enough comparisons and computational evaluation The main comparisons are only against OT-CFM with a small coupling context. They do not compare with existing large-scale coupling methods, such as offline Sinkhorn [1] and semidiscrete methods [2,3]. These experiments show that increasing the coupling context helps, but they do not show that the proposed method is better than strong baselines. Also, the reported coupling latency and share of per-step runtime do not show overall training efficiency. The authors need comparisons under the same total training time, or time to reach a target generation quality. This matters because practical scalability is a main motivation of the paper.
  2. Limited methodological novelty Minibatch OT pairing [4], entropic couplings, and the benefits of larger coupling sizes have been studied before. The proposed training procedure mostly combines these ideas with the existing FlashSinkhorn solver [5]. The online implementation is useful, but the paper does not clearly show a major methodological advance beyond this combination. The theoretical analysis therefore has to carry much of the novelty claim. However, it relies on population optima that can be reached in practice and only covers one-step Euler error, so it does not directly explain practical training and sampling.

References: [1] Stephen Y. Zhang et al., “On Fitting Flow Models with Large Sinkhorn Couplings,” TMLR, 2026.

[2] Alireza Mousavi-Hosseini et al., “Flow Matching with Semidiscrete Couplings,” ICLR, 2026.

[3] Lingkai Kong et al., “AlignFlow: Improving Flow-based Generative Models with Semi-Discrete Optimal Transport,” ICLR, 2026.

[4] Aram-Alexandre Pooladian et al., “Multisample Flow Matching: Straightening Flows with Minibatch Couplings,” ICML, 2023.

[5] Felix X.-F. Ye et al., “FlashSinkhorn: IO-Aware Entropic Optimal Transport on GPU,” ICML, 2026.

Questions for authors

Please see the weaknesses

3 comments
AuthorsOpenReview ↗

Author response (1/2)

We genuinely appreciate your time to provide these constructive and helpful suggestions. We address each point below:


W1.1: Comparison to existing large-scale coupling methods.

We agree and have added AlignFlow (ICLR 2026), with its semi-discrete OT map pre-computed over the full dataset. Under the same number of training steps, it achieves an adaptive FID of 7.72, improving over OT-CFM (7.89) and matching EOT-CFM-2048 (7.71). We provide training time details in our next response, where we also extend steps to compare them under training time.

We note that offline semi-discrete coupling offers a practical trade-off between precomputation cost and quality. For offline Sinkhorn coupling method, the reported implementation requires a distributed JAX to pre-compute full-scale couplings up to sixteen 80GB-memory GPUs, which is computationally infeasible for us. Moreover, different from online coupling methods, the authors of AlignFlow also acknowledge that current offline coupling formulation requires further adaptation for more complex text-conditioned generation, left as future work.

Therefore, to further demonstrate the practical advantages of our method, we have added text-to-image training over CUB dataset, with detailed training configs provided in section 5. We report FIDs below along with qualitative results in Figure 4.

NFE 25NFE 100Adaptive
OT-CFM-12811.0410.6110.27
EOT-CFM-409610.8810.4410.12
EOT-CFM-819210.4910.209.94

Beyond the improvement in generation quality, we find that larger coupling minibatches also improve text alignment. For example, in the third panel of Figure 4, the prompt is "bird with a long beak, blue eyes, and dark gray and black feathers", and both baselines incorrectly assign blue to the feathers, while ours with a larger coupling minibatch follows the proper feather color.


W1.2: The authors need comparisons under the same total training time, or time to reach a target generation quality.

We apologize for the omission. On CIFAR-100, OT-CFM takes 12.2 h, offline semi-discrete coupling takes 12.9 h, and our largest EOT-CFM-8192 takes 14.3 h. Our adaptive minibatch schedule (EOT-CFM-scheduled, detailed in our next response) achieves performance comparable to EOT-CFM-8192 in only 12.5 h. Shown in Figure 5, we train the other methods for additional steps to extend their training time to 14 h, and our one with larger minibatch still performs better under both same training time and same training steps. We also note that the most of the additional overhead comes from coupling computation, so with larger task or more complexed text-conditioned task, this overhead will account for a smaller fraction of total training time, as flow optimization process takes most of the training time (e.g., on ImageNet, OT-CFM takes 55.2 h with 13.26 FID, EOT-CFM-scheduled takes 56 h with 12.96 FID, and EOT-CFM-10K takes 58.8 h with 12.94 FID). We have also added this to the updated manuscript.

We note that FlashSinkhorn provides a practical starting point for efficient torch-based EOT and still continues to evolve through updates and GPU kernel optimizations (and its attention-style formulation offers future sustainable improvement to incorporate advances in FlashAttention as well), so we believe EOT-CFM would also further benefit from both sources of progress in future efficient-GPU EOT works.

AuthorsOpenReview ↗

Author response (2/2)

W2.1: The online implementation is useful, but the paper does not clearly show a major methodological advance beyond this combination

We discuss how the current results can provide further insights into the methodology. The upper bound in Theorem 4.15 suggests that larger minibatches can provide greater gradient variance reduction during earlier training iterations when model Jacobian norm is large (with further explanations deferred to Appendix B). Consistent with this theoretical insight, Figure 2 empirically shows that the variance reduction achieved by larger coupling minibatches is most pronounced during the first quarter of training.

Since we can now meaningfully scale the minibatch, a natural question is whether we can apply a scheduled adaptive coupling minibatch strategy, i.e., starting with a large minibatch to stabilize early training, then gradually reducing it as training becomes more stable, which keeps the early variance-reduction benefits while further saving training cost.

To justify this, we apply a minibatch scheduling strategy that starts with a large coupling minibatch of 8192 to stabilize early training, then gradually reduces it to 4096 and eventually to 2048. For unconditional generation on CIFAR-100, EOT-CFM-scheduled achieves 7.29 FID, comparable to 7.25 FID with EOT-CFM-8192, while requiring only 12.5 hours of training, approaching the 12.2 hours of OT-CFM; For text-conditioned generation, EOT-CFM-scheduled achieves 9.98 FID in 18.5 hours, compared to 9.94 FID in 20.2 hours for EOT-CFM-8192 and 10.27 FID in 18.1 hours for OT-CFM. These results suggest that minibatch scheduling retains the benefits of larger coupling minibatches while bringing training costs closer to those of OT-CFM.

We hope these findings further strengthen the connection between our theoretical results and their empirical significance.

W2.2: reliance on population optima & only covers one-step Euler error

Sorry for any confusion, we clarify that optimality is a common setup in FM analysis, and we have added discussion of more prior works under this setup in line 203-208. The most important reason is that optimality removes the variance contribution from imperfect fitting of the conditional-mean velocity, providing a common basis for comparing gradient noise under different coupling strategies.

We believe the most valuable significance for our one-step Euler error is to establish a provable framework for how coupling quality affects the generated distribution. As discussed in section 5.1 of multi-sample FM (Pooladian et al., 2023), a coupling with correct marginals doesn't automatically produce the correct distribution after a conditional-mean Euler update. Our result makes this connection explicit by relating the entropic OT objective to a Wasserstein error bound, thereby bridging the theoretical gap between coupling quality and generation performance.


In the end, we again sincerely appreciate your time in providing these useful feedbacks, which makes our evaluation and contribution better. We hope that our responses have addressed your concerns, and we look forward to an open-minded discussion if any such concerns remain.

Sincerely,

Authors of Paper 187

Hi reviewer Q4Vn,

As the rebuttal period comes to the end, could you please check our response to see if the final concern is resolved? Thanks again for your time.

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

This paper investigates the scalability of online optimal transport (OT) coupling for flow matching. By leveraging FlashSinkhorn, the authors enable substantially larger minibatch couplings while maintaining manageable computational overhead. The paper further provides a comprehensive theoretical analysis of how OT coupling affects stochastic-gradient variance and integration error, including the influence of entropic regularization and minibatch size. Experiments on synthetic distributions, single-cell datasets, CIFAR-100, and ImageNet-64 support the theoretical findings.

Critical strengths
  1. Solid and comprehensive theoretical analysis. The main strength of this paper lies in its theoretical contributions. The authors provide a systematic analysis of stochastic-gradient variance under independent, optimal, and entropically regularized couplings. In particular, the analysis establishes connections between transport cost, conditional velocity uncertainty, and full stochastic-gradient variance. The theoretical treatment of entropic regularization and minibatch scaling is also valuable.

  2. Practical and efficient scaling of online OT coupling. The proposed approach enables substantially larger OT coupling contexts without introducing prohibitive memory or computational overhead. The experiments generally demonstrate improved training stability and generation quality as the coupling minibatch increases.

Critical weaknesses
  1. Limited scalability to conditional generation.

My primary concern is the applicability of the proposed approach to conditional generation, particularly text-conditioned image and video generation. The method relies on constructing a sufficiently large source-target coupling context. However, in conditional generation, the available target samples associated with an individual condition can be extremely limited. For instance, text-to-image datasets often provide only one or a few reference images for a particular prompt. Consequently, constructing thousands of candidate pairs under the same condition may be impractical.

Moreover, the empirical improvements exhibit diminishing returns as the coupling minibatch increases. For example, on ImageNet-64, increasing the context from 8192 to 10240 only improves the adaptive-sampling FID from 13.01 to 12.94. Although this does not invalidate the method, it raises questions about the practical benefit of scaling to even larger coupling contexts. Ideally, experiments on a class-conditional or text-conditional generation task would help establish the practical applicability of the approach.

  1. Insufficient comparisons with independent coupling and existing variance-reduction methods.

Although the paper provides theoretical comparisons with independent coupling and reports its gradient variance in Figure 2, the main image-generation results in Table 2 do not include the generation performance of standard independent-coupling CFM.

This is an important omission because the paper's central motivation is that OT coupling reduces stochastic-gradient variance and thereby improves optimization and generation quality. However, the practical significance of this improvement remains unclear without a direct comparison against a well-optimized independent-coupling baseline. In particular, since EMA can mitigate the effects of noisy parameter updates, it would be interesting to examine how much of the theoretical variance-reduction advantage translates into improved generation quality under this standard training practice.

Furthermore, alternative approaches explicitly target variance reduction without requiring large-scale OT coupling. Representative examples include Stable Velocity Matching (StableVM) from Stable Velocity: A Variance Perspective on Flow Matching, This work is relevant to the broader discussion of variance reduction. I believe comparisons with independent CFM and representative FM-specific variance-reduction methods, under matched architectures and computational budgets, would substantially strengthen the empirical contribution.

Questions for authors

Same to the weaknesses

5 comments
AuthorsOpenReview ↗

Author response

We sincerely appreciate your insightful and valuable input. We answer them one-by-one below:


W1: experiments on text-conditional generation

Great point. We add text-conditioned training experiments on CUB dataset, with detailed training configs at the bottom of page 9. We provide both quantitative FID results and qualitative examples in Figure 4.

NFE 25NFE 100Adaptive
I-CFM11.8811.0210.58
OT-CFM-12811.0410.6110.27
EOT-CFM-409610.8810.4410.12
EOT-CFM-819210.4910.209.94

Beyond generation quality, we find that larger minibatch context also improves text alignment.

For example, in the third panel of Figure 4, the prompt is "bird with a long beak, blue eyes, and dark gray and black feathers". However, both baselines incorrectly assign blue to the feathers, while ours with scaled coupling minibatch follows the proper feather color, which shows that larger coupling minibatch would provide a more consistent conditional training signal that helps associate text attributes with their intended visual components.

Thanks again for bringing out this important point.


W2.1: comparison to independent CFM

We clarify that the original OT-CFM paper has already compared to I-CFM on both single-cell and image generation task (their Table 4 and 5), so we take OT-CFM as better baseline. Still, to make it explicit, for un-conditional task, I-CFM's adaptive FID is 7.96 on CIFAR-100 (comparing to 7.89 for OT-CFM) and 13.44 (comparing to 13.26 for OT-CFM) on ImageNet under EMA checkpoints. For conditional task, we report it above in the response towards W1. Regarding the training budget, we note that OT-CFM has nearly the same training speed as I-CFM, which is reported in OT-CFM paper with less than 1% computational overhead. Still, to further justify our empirical contributions, we extend the training of the better baselines to compare with EOT-CFM under the same training budget. As shown in Figure 5, our one with larger minibatches outperforms the baselines under both matched training time and matched training steps.

We note that FlashSinkhorn provides a practical starting point for efficient torch-based EOT and still continues to evolve through updates and GPU kernel optimizations (and its attention-style formulation offers future sustainable improvement to incorporate advances in FlashAttention as well), so we believe EOT-CFM would also benefit from both sources of progress in future efficient-GPU EOT works.


W2.2: broader discussion of variance reduction

Thanks for pointing out this. We have cited StableVM and added a discussion in the main text, along with a broader discussion of stable FM in the appendix. Our work focuses on improving the stability of minibatch FM through training pair construction, which is complementary to approaches that improve stability by refining supervision estimation. We believe these two directions open up promising opportunities for future work to explore how they can be combined to further improve FM stability both theoretically and empirically.


Again, thank you for your close reading and insightful feedback (especially for pointing out T2I experiments), which we believe has significantly strengthened our empirical results. We hope that our responses have addressed your concerns, and we look forward to an open-minded discussion if any such concerns remain.

Sincerely,

Authors of Paper 187

Thank you for the additional experiments. The new results substantially address my concerns and I will defend my positive score.

However, I would appreciate clarification on how OT coupling is constructed for text-conditioned generation. Is the coupling computed globally across images with different captions, or separately within each condition? If global coupling is used, the matched source noise may become correlated with the text condition, resulting in pOT(z∣c)≠p0(z)p_{\mathrm{OT}}(z\mid c)\neq p_0(z). This train-test mismatch has been studied in The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation (https://arxiv.org/abs/2503.10636). Could the authors clarify whether this issue is relevant to their method and, if so, whether entropic regularization mitigates it?

Many thanks for your prompt reply, we share our thoughts on this.

First, we use the global coupling, so the curse of condition also applied here. C$^2$OT use a direct way to address the issue of noise–condition dependence, while entropic coupling provides a different mechanism by penalizing deviations from independent coupling.

Recall that entropic coupling solves

πε=arg⁡min⁡π∈U(a,b)⟨C,π⟩+εKL(π|a⊗b)\pi_\varepsilon=\arg\min_{\pi\in U(a,b)}\langle C,\pi\rangle+\varepsilon\mathrm{KL}\left(\pi\middle|a\otimes b\right),

where aa and bb are the empirical marginal weights of the source noise and target image–caption pairs, so the KL term actually encourages noise-target independence and would consequently alleviate noise–condition dependence.

We will revise the manuscript to clarify our global coupling setup and provide more discussion on this question. Indeed, condition-aware coupling, such as C$^2$OT, can in principle be combined with entropic OT, and we believe it's an interesting direction for future work to investigate to what extent entropic regularization mitigates noise-condition dependence (both theoretically and empirically) and how the two interact when combined.

Thanks for your insights on this interesting question. We also again appreciate your time spent to help us improve the work.


Sincerely,

Authors of Paper 187

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thanks for your time: author response & updated manuscript

Dear reviewers,

Thank you all for your time in providing the feedback. We have provided detailed responses and updated the manuscript, with major changes in green. We again appreciate your time to check our response.


Sincerely,

Authors of Paper 187

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.