acceptodds
Under review as a conference paper at ICLR 2027

When Can Diffusion Denoisers Share a Representation Across Noise Levels?

Abstract

Diffusion models solve denoising problems across noise levels, yet typically reuse a single time-conditioned network. Motivated in part by latent diffusion, we ask when the data distribution itself justifies sharing a common representation across noise levels. For Gaussian diffusion, we establish an exact characterization: if the Bayes-denoiser Jacobian is diagonalizable in the same fixed frame at every input at just one positive noise level, then the clean distribution admits an affine orthogonal independent-component representation. Conversely, this structure guarantees a common coordinatewise Bayes-denoising frame across all noise levels. Beyond exact independence, we show that total correlation equals both the noiseintegrated excess Bayes risk of coordinatewise denoising and the generative KL discrepancy of an ideal product model. We distinguish this factorization cost from the additional cost of sharing a frame across noise levels. We derive sufficient finite-sample learning bounds for shared and untied frames, and show why approximate positive-noise Bayes-denoiser separability need not imply approximate clean independence in a prescribed frame. Experiments with independently trained denoisers and learned VAE-family and CIFAR-10 representations support these predictions, including the finite-sample crossover between shared and noisespecific representations.

Primary area:
learning theory
Keywords:
Diffusion Models, Shared Representations, Bayes-Optimal Denoising, Independent Component Analysis, Total Correlation

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

This paper studies when the data distribution supports a shared denoising representation across Gaussian noise levels. It proves that a fixed frame diagonalizing the Bayes-denoiser Jacobian at every input at one positive noise level is equivalent to an affine orthogonal independent-component structure, enabling coordinatewise Bayes denoising in the same frame at all noise levels. It connects total correlation to integrated denoising regret and ideal product-model KL discrepancy, separates factorization and sharing costs, and derives finite-sample learning bounds. Experiments with independently trained denoisers, VAE-family representations, and CIFAR-10 latents support common-frame transfer and a finite-sample crossover between shared and noise-specific representations.

Critical strengths
  • The paper offers an interesting ICA-based perspective on when denoisers can share a representation across noise levels, providing useful insight into the structure that enables such sharing.
  • The paper distinguishes the cost of coordinatewise denoising from the additional cost of sharing a frame, helping explain when sharing can be beneficial and when it introduces bias.
Critical weaknesses
  • Limited validation on complex image data. CIFAR-10 images are only 32 × 32 pixels and are further compressed into 32-dimensional latents. Coordinate-shuffling controls also reduce dependence to better match the theoretical assumptions. While useful for testing the mechanism, these experiments leave applicability to unmodified, high-resolution images with richer structure and fine details unclear.

  • Uncertain reliability of a fixed frame across inputs and noise levels. Empirical frames are estimated from finite samples at one noise level, whereas the exact theorem requires the Jacobian condition to hold at every input. Although held-out transfer tests and controlled tying-bias experiments provide supporting evidence, they do not establish reliable frame recovery or suitability throughout denoising on complex image distributions.

  • Limited connection to practical architectures. Fixed orthogonal, coordinatewise denoising is considerably more restrictive than the nonlinear, interacting features used by practical diffusion networks. Comparisons with time-conditioned or noise-specialized architectures are absent, leaving the implications for practical sharing strategies, generation quality, efficiency, and robustness unclear.

  • Insufficiently demonstrated added explanatory value. Although the paper discusses related theory, the experiments primarily validate its own predictions. A clearer comparison with relevant ICA or covariance-based perspectives would help establish what additional insight the framework provides into cross-noise frame transfer and the benefits of sharing.

Questions for authors
  • Applicability to complex data: What do the results on the original, unshuffled latents establish about the usefulness of a shared frame? Which findings support extending this interpretation beyond low-dimensional representations of simple images, and which remain untested?

  • Computational implications: Does the shared-frame structure imply computational savings beyond reduced frame-estimation complexity? Please clarify whether lower model size, training cost, or inference cost follows from the analysis, or remains a potential future application.

  • Source-level sensitivity: Appendix I.5 evaluates multiple source noise levels for dSprites. How sensitive are the CIFAR-10 results to the choice of the source level, σ₀ = 0.6? In particular, is transfer to low-noise stages consistent across different source levels?

  • Connection to practical sharing: Which aspects of representation sharing in standard time-conditioned diffusion networks are captured by the fixed orthogonal, coordinatewise model? What insight does the analysis provide when useful denoising requires cross-coordinate interactions?

4 comments

We thank the reviewer for their careful reading and constructive feedback. Regarding presentation, we have revised the manuscript to improve clarity and organization, and invite the reviewer to consult the updated PDF.

Regarding applicability to complex data, modern high-resolution diffusion models often operate in learned latent spaces rather than directly in pixel space. Our paper characterizes when such latent distributions support a shared representation for the Bayes-optimal denoiser, the target of squared-error diffusion training. Our theory neither assumes linear encoders nor restricts the latent dimension. Importantly, coordinate shuffling is only an experimental control. We also evaluate the original dependent CIFAR-10 latents, whose fixed-frame residual is substantially larger than that of the shuffled controls (0.05356 versus 0.002372), yet whose additional cross-noise transfer error remains relatively small (0.001336). This suggests partial geometric stability even without approximate independence.

Regarding frame reliability, we have added Appendix L, which establishes quantitative approximate one-noise transfer guarantees using average Bayes-Jacobian defects, including blockwise extensions and a conditional held-out validation certificate for estimated frames. These results address imperfect diagonalization and finite-sample evaluation under explicit assumptions.

Regarding practical architectures, our denoisers are not necessarily linear: the coordinatewise functions can be arbitrarily nonlinear and noise-dependent. Their lack of cross-coordinate interactions is precisely the simplification justified by our theory. Moreover, Appendix K extends the characterization to independent blocks, permitting arbitrary nonlinear interactions within blocks. Our regret and tying-cost analysis quantifies the consequences of approximate sharing. Sections 6.1–6.2 investigate independently trained denoisers and shared-versus-untied frame estimation; our CIFAR experiments additionally examine time-conditioned networks.

Regarding computational implications, Section 5 establishes finite-sample benefits of estimating a shared frame, while Appendix K.2 demonstrates explicit computational savings for exact blockwise Gaussian-mixture denoising, including the possibility of independent or parallel block evaluation. We do not claim general reductions in neural-network size, training cost, or inference latency.

Regarding explanatory value beyond ICA and covariance, covariance eigenvectors remain unchanged under isotropic Gaussian corruption for every distribution, yet this does not imply shared Bayes-denoising coordinates (Appendix F.1). Beyond classical ICA, our framework establishes the one-noise Bayes-Jacobian characterization, its cross-noise consequences, and quantitative results for approximate sharing, tying bias, and independent blocks.

Regarding source-level sensitivity, σ0=0.6\sigma_0=0.6 was chosen as a representative level rather than optimized for transfer. The exact characterization applies at any positive noise level, and Appendix I.5 investigates multiple sources on dSprites. Additional CIFAR-10 source-level comparisons, particularly high-to-low noise transfer, remain valuable.

Finally, our paper is primarily theoretical: we characterize when representation sharing is justified by the Bayes-optimal target itself, rather than proposing a particular neural architecture or claiming improved end-to-end generation. Our experiments test the underlying mechanisms; validating practical architectural gains on larger spatial latents remains future work.

Thank you for the clarifications and for adding Appendix L. The new analysis establishes a conditional transfer guarantee for targets at or above the source noise level. My concerns about practical applicability remain.

  • Transfer to lower noise levels. The CIFAR-10 experiment fits the frame at σ₀ = 0.6 and reports a mean excess Jacobian residual of 0.001336 across three target levels. Reporting the excess residual for each target across several source noise levels would clarify whether it stays small when transferring to lower-noise targets on unmodified latents.
  • Connection to denoising performance. The CIFAR-10 results measure the geometry of learned network Jacobians. Applying Corollary L.7 requires an independent bound on their error relative to the Bayes Jacobians. A controlled comparison with a standard time-conditioned denoiser on the same latents would help establish whether explicitly sharing a frame can achieve comparable denoising quality with fewer training samples.

I maintain my original rating of 2 because the practical benefit of explicitly sharing a frame in learned diffusion models remains insufficiently demonstrated.

We thank the reviewer for their thoughtful follow-up and for acknowledging the new theoretical guarantees in Appendix L. We appreciate the opportunity to clarify two important points.

Regarding lower-noise transfer, our CIFAR-10 experiment already evaluates a frame learned at σ0=0.6\sigma_0=0.6 on the original, unshuffled latents at σ∈{0.15,0.30,1.20}\sigma\in\{0.15,0.30,1.20\}, including two substantially lower-noise targets. The reported excess residual of 0.001336±0.0001020.001336\pm0.000102 is averaged over these three targets, relative to independently fitted target-level frames. Thus, the existing results already provide evidence of lower-noise transfer on dependent real-image latents. We agree that a target-specific breakdown across multiple source levels would further illuminate its robustness. Importantly, the exact theorem applies to any positive source level, while Appendix L provides complementary quantitative guarantees under explicit assumptions.

Regarding denoising performance, we would emphasize a distinction central to our contribution: sharing a representation does not require eliminating all cross-coordinate interactions. On the original CIFAR-10 latents, the fixed-frame residual is 0.053560.05356, yet the additional cross-noise transfer residual averages only 0.0013360.001336. This suggests that a representation can remain relatively stable across noise levels even when substantial coordinate interactions persist. Our theory distinguishes these effects through factorization regret and tying cost, and Appendix K extends the framework to nonlinear, interacting denoisers within shared blocks.

Moreover, Section 6.2 directly evaluates denoising risk and demonstrates the finite-sample sharing–tying tradeoff. The CIFAR-10 experiments complement this by studying representation geometry in fully interacting, time-conditioned networks. We appreciate the proposed architectural comparison as a further test of sample efficiency.

Our central contribution is to establish when and why cross-noise representation sharing is supported by the Bayes-optimal denoising target itself, providing a mathematical foundation for investigating practical sharing strategies.

Thank you for the additional explanations. The results support cross-noise representation transfer in the studied settings, but without comparisons to standard time-conditioned diffusion models or other strong practical designs, it remains unclear whether this translates into practical benefits. I therefore maintain my current rating

Reviewer_V4vYOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper asks when the data distribution justifies sharing one coordinate frame across noise levels in diffusion denoising.

  • Theorem 3.1 shows that the optimal denoiser at one noise level acts coordinatewise in a fixed frame if and only if the data are an orthogonal linear mixture of independent sources.
  • Corollary 3.2 shows that if the optimal denoiser acts coordinatewise in a fixed frame at one noise level, the same frame is optimal at all noise levels.
  • Theorem 4.1 quantifies the loss of coordinatewise denoising, showing that its integral over noise levels equals the total correlation of the data.
  • Proposition 4.2 shows that the converse of Theorem 3.1 does not hold approximately: a nearly coordinatewise denoiser does not imply nearly independent data.
  • Section 5 bounds the finite-sample excess risk of learning the frame in a specific model, showing that one shared frame needs fewer samples than separate frames per noise level.
  • Section 6 empirically verifies these predictions on synthetic data and learned latent representations.
Critical strengths
  • Rigorous analysis under the assumption. Under its stated assumptions, the paper gives clean and correct results.
  • Experiments consistent with the theory. The experiments are well designed to test the theory, and their results are consistent with its predictions.
Critical weaknesses
  • Assumptions far from practical settings. The exact result assumes that the data are an orthogonal linear mixture of independent sources, which can be sampled without diffusion. Beyond this case, the connection to realistic data remains weak. The paper equates the integrated coordinatewise denoising loss with total correlation, but both quantities are intractable for realistic data, and the paper gives no bound on this loss in terms of the structure of the data. As Proposition 4.2 acknowledges, an approximately separable denoiser cannot certify approximately independent data, which limits the practical use of the main result.
  • Experiments do not reach practical diffusion settings. Most experiments use low-dimensional synthetic data with planted structure, and the only real-image experiment uses a small autoencoder latent space. It would strengthen the paper to measure how far widely used settings, such as Stable Diffusion latents or pixel-space CIFAR-10, are from the shared-frame regime, and whether this informs model design.
Questions for authors
  1. How can we leverage the paper's results? Can they guide the design of practical denoisers, for example in choosing the latent representation or deciding how to share parameters across noise levels?
  2. mtm_t first appears in Eq. (3) without a proper definition.
  3. Section 4 introduces new notation (ZZ, ZtZ_t, VV, O(k)O(k)) that appears to duplicate XX, YtY_t, and UU from Section 3 without explanation. What are their definitions?
4 comments

We thank the reviewer for their careful review and valuable feedback. Regarding the presentation, we have substantially revised the manuscript to improve its clarity and readability. The reviewer can find these improvements in the updated PDF.

Regarding the first weakness, we are not sure whether our main contribution was conveyed clearly enough. The paper does not assume that "the data are an orthogonal linear mixture of independent sources, which can be sampled without diffusion." Rather, we start from the Bayes-optimal denoiser, the function that diffusion networks aim to learn, and ask why such a function should share a representation across noise levels. What we prove is that a particular structure in the clean distribution is equivalent to a shared denoising frame, and remarkably, this can be characterized from the Bayes-optimal denoiser at just one positive noise level.

We agree that the connection to realistic data is important. To address this concern, we have extended our characterization to independent orthogonal blocks, allowing arbitrary dependence within each block (Theorem K.1), which considerably broadens the class of distributions covered by our theory.

Regarding intractability and Proposition 4.2, computational tractability is not essential to the theoretical question we investigate. Moreover, diffusion is commonly performed in learned latent spaces, and our experiments with VAE-family representations show that the predicted structure can arise approximately in such spaces. Importantly, the certificate discussed in Proposition 4.2 is not our main practical takeaway. What matters for diffusion is the simplification and cross-noise representation sharing in the Bayes-optimal denoiser, together with quantitative guarantees on the loss incurred when independence holds only approximately. Proposition 4.2 limits the reverse inference from approximate separability to independence, not these forward guarantees.

Regarding practical diffusion experiments, we agree that we do not demonstrate improved end-to-end generation in a large-scale diffusion model. However, our paper is primarily theoretical: its goal is to explain why representation sharing in diffusion models can be justified by the data distribution, and what properties of a learned latent space can support it. Our experiments are designed to investigate these theoretical predictions.

Regarding Question 1, yes, our results suggest using latent-learning methods, such as dependence-regularized VAEs, to encourage representations with the desired structure. In such representations, a frame learned at one noise level can transfer to others, allowing observations from different noise levels to contribute to learning the same representation. This can reduce sample complexity for a desired estimation accuracy, as motivated by our finite-sample analysis and demonstrated in our controlled experiments. Specifically, Figure 1 (Section 6.1) demonstrates cross-noise frame transfer, while Figure 2 (top panels, Section 6.2; details in Appendix G) shows that sharing substantially reduces excess denoising risk compared with learning separate frames under the same total training budget.

Regarding Question 2, the reviewer is correct. We have fixed the missing definition in the revised manuscript.

Regarding Question 3, these notations are not duplicates. As we move from the exact characterization to approximate sharing, motivated in part by latent diffusion, we use ZZ, ZtZ_t, and V∈O(k)V\in O(k) to work directly in the latent representation space, rather than the original ambient space described by XX, YtY_t, and UU. We agree that this transition was not sufficiently clear, and have explicitly explained it in the revised Section 4.1.

We hope that the revised manuscript and these clarifications better convey the motivation and contributions of our work.

Reviewer_V4vYOpenReview ↗

Response to Authors

I thank the authors for the detailed responses and the revised manuscript. I summarize the status of each point below.

CW1: Partly addressed.

  • The main theoretical result of the paper is that the Bayes-optimal denoiser shares one coordinate frame across all noise levels if and only if the data have independent components in that frame, and that checking a single noise level suffices. Whichever side is regarded as the assumption, this setting is far from realistic data.
  • If a latent is nearly independent, the loss of coordinatewise denoising is small, but there is no way to check whether a realistic latent is nearly independent: the denoiser cannot reveal it (Proposition 4.2, C.1 in the revision), and total correlation cannot be evaluated.
  • The block extension partly relaxes this setting.

CW2: Not addressed. My concern is not the scale of the data. Even on a small dataset, the experiments should cover the setting that diffusion models actually face. The experiments are consistent with the theory, but the favorable regime is reached only with coordinate-shuffled CIFAR-10 latents, which are not encodings of real images.

Q1: Not addressed. The response only suggests dependence-regularized latents with a shared frame, but does not show how the findings can be used to construct such a latent space for generation, for example how to set the regularization strength, which may ease denoising but harm reconstruction.

Q2 and Q3: Addressed. I thank the authors for defining mtm_t and for clarifying the transition to the latent notation in Section 4.1.

We thank the reviewer for the thoughtful follow-up and the opportunity to further clarify our contributions.

Regarding CW1, an important contribution of our work is understanding representation sharing beyond exact independence. Theorem K.1 extends our characterization to independent blocks with arbitrary within-block dependence, while Eq. (11) distinguishes the cost of coordinatewise approximation from the additional cost of sharing across noise levels. Thus, even when coordinates are dependent, a common representation can remain useful. Furthermore, Appendix L establishes quantitative guarantees relating one-noise Bayes-Jacobian defects to higher-noise denoising regret under stated regularity conditions.

Regarding CW2, our experiments also evaluate the original, unshuffled CIFAR-10 latents, which are genuine representations of real images. Interestingly, despite their non-negligible fixed-frame residual (0.05356), the additional cross-noise transfer residual is only 0.001336 (Appendix I.13–I.14). This illustrates that useful cross-noise geometric transfer can occur even beyond the nearly independent regime. The shuffled experiments complement this observation by isolating the effect of dependence.

Regarding Q1, our theory suggests a concrete approach: learning latent representations with coordinatewise or blockwise dependence regularization, while selecting its strength based on reconstruction quality and held-out denoising performance. Theorems 4.1 and K.3 quantify the denoising cost associated with dependence, while Eq. (11) and Section 5 provide a framework for balancing approximation cost against the statistical benefits of sharing.

We hope these clarifications highlight how our results provide both a theoretical explanation for cross-noise representation sharing and practical guidance for designing and evaluating structured diffusion denoisers.

I thank the authors for their further response. However, my main concerns remain unresolved, and I therefore maintain my original score. My concern is not with the correctness of the theorem, but with what we learn from it: it remains unclear whether the result provides a nontrivial insight into diffusion denoising or offers useful guidance for practical generative modeling. The restrictive assumptions and the limited experimental validation leave this central issue unresolved.

  • While I find the proposed theorem sound, its applicability is limited to idealized settings. The response does not establish how its conclusions extend to practical generative modeling.
  • The experiments do not demonstrate the claimed implications in a setting that supports end-to-end generation. Comparing shuffled and unshuffled CIFAR-10 latents provides a controlled comparison, but the unshuffled baseline alone does not establish the practical utility of the proposed approach.
  • Encouraging a more linear latent geometry may make denoising easier, but it remains unclear how this affects the representation's ability to capture complex data structure. An investigation of the trade-off between representational expressivity and ease of denoising would strengthen the paper. This trade-off, however, is neither evaluated nor explored.
Reviewer_Zq4LOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors study when the same coordinate system can be used for optimal denoising across different noise levels. They show that if the optimal denoiser acts independently on each coordinate at one positive noise level, then the clean data have independent components in that coordinate system, and the same structure holds at every noise level.

Critical strengths

The authors propose interesting and relevant problem that could potentially improve the vanilla diffusion models for certain applications. They also use the latents of CIFAR-10 dataset, which I think are complex enough to study such phenomena.

Critical weaknesses

The manuscript is difficult to follow in its current form. The mathematical notation needs clarification and greater consistency. For example, (D) denotes both a diagonal matrix and a derivative in the same equation (line 52), while (k) appears without explanation (line 51). The Jacobian notation (J) is introduced in line 184, although (Df(x)) is used several times earlie which creates confusion. Similarly, (m_t) in equation (3) is not defined.

The presentation also needs improvement. The experimental results would be easier to read if the numerical values were organized into tables rather than listed throughout the text. The explanation of the method also needs more intuition to help readers understand the mathematical results and their significance.

Taken together, these issues prevent a clear understanding of the paper in its current form.

Questions for authors

What are the potential practical applications of the proposed method, and in which settings could it offer an advantage over standard diffusion setting?

3 comments

We thank the reviewer for their feedback. We have substantially improved the paper's presentation and addressed the specific concerns regarding notation, definitions, intuition, and experimental organization in the revised PDF.

Our starting point is the Bayes-optimal denoiser, which diffusion networks aim to approximate. It maps a noisy observation and its noise level to the optimal denoised estimate. Although diffusion architectures typically share parameters across noise levels, why should the optimal denoising function itself admit a shared representation?

Our paper provides a mathematical answer: under specific structures in the clean distribution, the Bayes-optimal denoiser admits a common representation across all noise levels. We also establish the converse from a single positive noise level, extend the characterization to independent blocks, and quantify the costs of approximate factorization and sharing.

Importantly, these assumptions are relevant to latent diffusion, where the data distribution is shaped by a learned encoder. Our results explain why appropriately structured latent spaces can simplify optimal denoising and potentially reduce learning complexity.

Regarding practical advantages, Section 5 provides finite-sample bounds for the statistical benefit of sharing, while our experiments demonstrate both its advantages and limitations.

We would also appreciate clarification of any specific mathematical soundness concerns beyond the presentation issues raised in the review.

I thank the authors for their response and the revisions to the manuscript. The readability of the paper has indeed improved. Although I still find it quite dense and difficult to follow, I appreciate that the notation is now clearer and the presentation is more logical.

However, I believe that my question about the practical relevance of the theoretical results is still not addressed. In particular, it is unclear to me why characterizing the conditions under which the optimal denoiser can use a shared representation is important in practice. Modern diffusion models already share most of their parameters across timesteps while using timestep embeddings to adapt their behavior to different noise levels. This allows the model to learn how much timestep-dependent information is needed. Therefore, I am not sure how the proposed theoretical results could help improve the design or training of such models. Additionally, the assumptions about the data distribution seem rather restrictive. In practice, latent representations often contain strong dependencies between dimensions, including those learned by VAEs in latent diffusion models. Although the authors also consider approximate factorization, it is still unclear how well these assumptions hold for realistic latent representations and whether the theoretical results remain useful in such settings.

Overall, I appreciate the idea behind the paper and its theoretical contributions. However, despite the improvements in the revised version, I still find the manuscript difficult to follow, and I believe that both the clarity of the presentation and the practical motivation need further improvement to meet ICLR standards.

We thank the reviewer for revisiting the manuscript and acknowledging the improvements in its presentation. We would like to clarify the practical significance of our results, as we believe an important distinction remains.

We agree that modern diffusion networks already share parameters across timesteps and use time conditioning to adapt their behavior. However, this architectural flexibility does not explain whether the optimal denoising functions actually require different representations at different noise levels. A time-conditioned network can implement entirely different coordinate interactions at each timestep, even when its parameters are shared.

Our paper addresses precisely this question. Theorem 3.1 establishes an equivalence between a one-noise Bayes-Jacobian structure and independent components in the clean distribution, implying that a single orthogonal frame is sufficient across all noise levels. This is stronger than observing that shared networks can learn to denoise at multiple timesteps: it identifies when a restricted, shared representation is mathematically sufficient without sacrificing Bayes optimality.

The practical implication is a principled inductive bias: rather than allowing arbitrary timestep-dependent representations, one can reuse a common frame while retaining noise-dependent denoising functions. When appropriate, this can reduce representation complexity. Section 5 provides finite-sample evidence for this statistical advantage in a controlled model, without claiming superiority over general-purpose diffusion architectures.

We also agree that exact coordinate independence is restrictive. This is why our analysis goes beyond that setting. The independent-block extension permits arbitrary dependencies within blocks, while Theorems 4.1 and K.3 quantify the denoising cost of approximate factorization without assuming independence. Our tying-cost analysis additionally identifies when enforcing a common frame can be detrimental. Appendix L provides conditional guarantees for approximate one-noise-to-higher-noise transfer.

Thus, our contribution is not a claim that existing diffusion models are incorrectly designed or that realistic latents are exactly independent. Rather, it establishes when structured sharing is justified, what is lost when its assumptions hold approximately, and why such structure may be statistically advantageous. We see this as a theoretical foundation for designing and evaluating structured latent representations, rather than an immediately deployable replacement for existing architectures.

We appreciate the reviewer's recognition of the theoretical contributions and the improved presentation. We would welcome reconsideration of the assessment in light of these clarifications. In particular, if any specific mathematical soundness concerns remain, we would appreciate their identification so that we can address them directly.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your time and thoughtful reviews. With the author–reviewer discussion period ending on November 18 (AoE), please review the authors' rebuttal and share your updated assessment. Specifically, let us know whether your main concerns have been resolved, and if not, which points remain open.

Your input before the deadline is greatly appreciated!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.