acceptodds
Under review as a conference paper at ICLR 2027

Parallel Picard-Duhamel Integration for Fast Neuro-Spectral Architectures

Abstract

Neuro-spectral architectures (NeuSA) are a class of physics-informed neural networks (PINNs) that have proven effective for solving time-dependent PDEs in the frequency domain. However, training and inference often rely on costly numerical integration of the learned vector field, typically using sequential Runge–Kutta time stepping, which limits temporal parallelism. We introduce PaPi, an integration scheme inspired by weakly nonlinear flows that enables time-Parallel inference through exponential integration and a Duhamel–Picard scheme. We show that the Duhamel–Picard iteration converges exponentially for all flows, and provide explicit error bounds that depend on how close the flow is to linear. Experiments show that Papi accelerates training and inference by up to two orders of magnitude, achieving training times of under one second for low-dimensional PDE learning problems, while maintaining competitive predictive accuracy.

Primary area:
learning on time series and dynamical systems
Keywords:
Inference, Integration, Neuro-Spectral, Neural ODEs, Neural PDEs, Physics-Informed

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_rrmgOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes PaPi, a parallel-in-time integration method designed to accelerate ODE solving in neuro-spectral architectures. The key idea is to decompose the dynamics into linear and nonlinear components, and combine Picard–Duhamel integration with parallel scans to replace the conventional sequential RK4 solver. Experiments on multiple PDE benchmarks demonstrate substantial improvements in training efficiency while maintaining competitive predictive accuracy.

Critical strengths
  • Well-motivated problem. This work addresses the computational bottleneck of sequential RK4 integration by introducing a parallel-in-time approach, potentially improving both training and inference efficiency.
  • Principled formulation. The proposed method is mathematically well motivated, with convergence analysis and error bounds providing theoretical support.
  • Strong empirical results. PaPi achieves significant speedups across multiple PDE benchmarks while maintaining competitive predictive accuracy.
Critical weaknesses
  • Limited methodological novelty. The core components, including Picard iteration, Duhamel's principle, exponential integration, and parallel scans, are well-established techniques. The contribution appears to lie primarily in their combination and adaptation to NeuSA.
  • Insufficient ablation studies. The individual contributions of key design choices, such as warm-starting, temporal loss weighting, and different quadrature schemes, are not sufficiently isolated.
  • Limited generality. The method heavily relies on decomposing the dynamics into linear and nonlinear components, with an efficiently computable linear propagator. Its effectiveness for strongly nonlinear dynamics or systems without a tractable linear component remains unclear.
  • Limited inference evaluation. Although the paper emphasizes inference acceleration, comprehensive end-to-end inference latency comparisons with existing solvers are lacking.
Questions for authors

See weakness.

2 comments

We thank the reviewer for the constructive feedback, and appreciate that they have found the method to be principled and well motivated. We have clarified PaPi’s contributions and expanded the evaluation in the revised manuscript. Answers to each remark are below.

Weakness 1 (Limited methodological novelty):

We have added a more in-depth literature review, discussing in greater detail Picard methods for diffusion-model sampling and training, as well as related work in numerical analysis. We would argue that, although each of the integration method's components are themselves classical, PaPi's approach to the combination of exponential integration and parallel Picard iteration in the spectral domain are, to the best of our knowledge, unique. In particular, they differ from related work in how they compute the linear flow via the diagonalization of differential operators in the spectral basis, offloading any non-diagonal corrections to the nonlinear component. This makes it easy to compose, differentiate, and scale, as well as conducive to PDE applications with NeuSA.

Still, the reviewer's input has motivated us to sharpen our claim to novelty in the revised manuscript. PaPi's main contribution is to combine these pre-existing methods and extend them to scientific machine learning applications. In particular, PaPi's unique implementation allows it to fully leverage the massively parallel capabilities of modern ML, along with the natural linear-nonlinear splitting induced by PDEs in the spectral domain and NeuSA's formulation. This specialization lies behind the strong experimental results.

We have also emphasized the role of our second contribution, which are the procedures for the use of PaPi in training. The orders-of-magnitude speedups in the training of these Neural DEs are only possible due to the use of warm-starting, which to the best of our knowledge is new. We further emphasize this contribution by adding a full theoretical and experimental treatment of this feature, proving that it provides a reliable training strategy and showing experimentally that the induced gradients point in the right direction.

Weakness 2 (Insufficient ablation studies):

This is a reasonable concern that has been voiced by the other reviewers. We have added to the revised draft a substantial number of new ablations, including warm-starting, temporal loss weighting, and different integration methods.

Weakness 3 (Limited generality):

We believe the remark about strongly nonlinear systems is due to suboptimal presentation on our end. In the new version of the draft we clarify the discussion of the role of nonlinearity and state plainly that PaPi converges unconditionally for all Lipschitz systems; it simply converges faster when its Lipschitz constant is smaller. We have removed all references to epsilon (which in the first draft implied smallness) and state all results in terms only of the nonlinear factor’s Lipschitz constant. To illustrate this, we have included an additional experiment on the inverse problem for the Navier-Stokes equation, a strongly nonlinear PDE from fluid dynamics. PaPi still achieves state-of-the-art performance with substantially faster training times:

Modelλ1−1\lambda _1-1(λ2−0.01)/0.01(\lambda_2-0.01)/0.01rMSE uuTT (s)
PINN1.2e-37.6e-27.0e-3331
NeuSA1.9e-35.0e-28.7e-31147
PaPI6.5e-41.2e-25.6e-339.1

Moreover, we would argue that the existence of a linear/nonlinear decomposition is a reasonable assumption in PDE applications; a large portion of the PDEs of physical interest seem to be decomposable in these terms. Still, this is a valid concern, and we have addressed it more explicitly in the revised draft's section 4.2, where we describe general procedures to extract linear-nonlinear decompositions from fully nonlinear terms, making the method more broadly applicable.

Weakness 4 (Limited inference evaluation):

We believe these concerns are answered in the revised drafts expanded ablation section.

Misc. (Presentation)

We have used the additional space in the rebuttal to substantially clean up and clarify the writing, introducing more explanations where the previous draft was terse and including more figures. We hope the current draft is accessible and easier to read.

Thanks for the reply and my concern has been addressed. But I decided to keep my current score.

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

The paper introduces PaPi, a parallel-in-time integration and training procedure for Neuro-Spectral Architectures (NeuSA). It separates a tractable linear operator from a learned nonlinear correction. By doing so, it propagates the linear component analytically, and updates the entire trajectory through Picard–Duhamel iterations. Training reuses detached trajectories across optimizer steps and differentiates through a small number of new sweeps. Validation covers six forward PDE benchmarks and an inverse Burgers problem, which illustrates substantial reductions in optimization time with competitive accuracy.

Critical strengths
  1. A well-motivated solution to a consequential computational bottleneck. Exact propagation of the tractable linear component avoids reconstructing stiff linear dynamics through ordinary Picard iteration, while associative scans reduce the temporal dependency depth of each sweep to (O(\log N)). This combination fits the diagonal or block-diagonal operators used by NeuSA particularly well.

  2. Good empirical results. The evaluation covers several classical PDE problems. The proposed method achieves substantial speedups.

Critical weaknesses
  1. The contribution relative to closely related parallel integrators might be insufficiently established. The key idea of the paper is a combination of exact linear propagation and whole-trajectory iteration on the nonlinear part. Similar ideas have been used in (Kooij et al. 2015; Zhang et al. 2022). PaPi’s use of a known, structured linear operator and its integration into NeuSA training could constitute a useful contribution. However, the paper does not adequately position this adaptation against these close precedents or establish what substantive advance it provides. This limits the assessment of the method’s novelty and significance.

G. L. Kooij, et al, A Block Krylov Subspace Implementation of the Time-Parallel Paraexp Method and its Extension for Nonlinear Partial Differential Equations, 2015. Q. Zhang, et al, Fast Sampling of Diffusion Models with Exponential Integrator, 2022.

  1. A sequential exponential-integrator baseline is missing. Appendix E controls framework differences, but a sequential exponential integrator could also avoid RK4’s linear stability restriction. Comparing it with PaPi in JAX, using the same NeuSA model and linear/nonlinear split at comparable accuracy, would establish whether PaPi provides substantial benefits beyond adopting an existing exponential integration method.

  2. The computational advantage under strongly nonlinear dynamics is insufficiently characterized. PaPi’s efficiency depends on obtaining an accurate trajectory with relatively few Picard–Duhamel sweeps. A stronger nonlinear remainder or a longer time horizon may require more sweeps, potentially reducing the benefit of temporal parallelism.

Questions for authors
  • What is the proposed method’s novelty compared to the related works?
  • How do the required Picard sweeps and runtime vary with the strength of the nonlinear remainder and the time horizon? At matched solution accuracy, in which regimes does PaPi retain an advantage over a suitable sequential exponential integrator?
  • How is (\varepsilon) selected? With the PDE task fixed and the network retrained for each value, how does its choice affect training time, final accuracy, and the required number of Picard sweeps?

Typo: In line 194, sokution should be solution.

2 comments

We sincerely thank the reviewer for their remarks. We appreciate that they find the novel implementation aspects and results to be strong and answer their points below. Tables have been shortened to save characters. Full versions are in the draft.

Weakness 1/ Question 1:

We were not aware of the work by Kooij et al. and thank the reviewer for bringing it to our attention. It is an important precedent, and has been added to the literature review. Still, the implementations differ in two important points:

  • First, their approach is built for multi-processor pipelines where time is subdivided into parallel blocks, with combination within each block taking place sequentially. In contrast, PaPi parallelizes over all time-steps simultaneously, as made possible with modern ML hardware.

  • Second, and most important, their approach is based on spatial coordinates and finite differences, forcing the introduction of Krylov schemes for flow approximation. This additional machinery is an obstacle to differentiability and may have restricted their scalability beyond 1D problems. Our approach, on the other hand, takes place in the spectral domain. This sidesteps the problem via the diagonalization of differential operators in the spectral basis (and offloading any non-diagonal corrections to the nonlinear component). This makes PaPi accurate, simple to extend to 2D and 3D problems, and fully differentiable.

We believe the second method (Zhang et al.) is less related to our own. It is an example of sequential exponential integration for diffusion models, with the main objective being higher-accuracy for sequential steps. We have nevertheless included an in-depth comparison with it in the new draft.

In any case, we have taken the reviewer’s comments into consideration and have sharpened the paper’s claim to novelty in the newest draft. We position PaPi not as discovering iterative approximations to Duhamel’s formula, but as extending its application to the context of neural differential equations and scientific machine learning. This means we extend these methods to be differentiable and to fully leverage the massively parallel capabilities of modern ML, along with the natural linear-nonlinear splitting induced by PDEs in the spectral domain.

Additionally, we have emphasized in the new draft our contributions to the use of these methods in optimization. In particular, a large portion of the speedups obtained is only possible due to the use of the warm-starting scheme for training, which to the best of our knowledge is new. We further emphasize this point by adding a full theoretical and experimental treatment of this feature.

Weakness 2/ Question 2:

The remark on sequential exponential integrators is reasonable. PaPi offers two structural improvements on base NeuSA: exponential integration and parallelizable inference, as we emphasize in the new draft. It is therefore fair to compare it with a sequential implementation of exponential integration. In the new draft, we have added an additional ablation to the appendix, comparing PaPi with a sequential 4th-order Exponential Time-Differencing implementation of NeuSA:

ProblemPaPi (rMSE)NeuSA (rMSE)ETD4 (rMSE)PaPi(s)NeuSA (s)ETD4 (s)
3D Layers1.8e-2±1.1e-21.8\text{e-}2\pm1.1\text{e-}27.6e-3±7.7e-47.6\text{e-}3\pm7.7\text{e-}44.8e-3±4.1e-44.8\text{e-}3\pm4.1\text{e-}4251577424
Sine-Gordon5.6e-4±7.0e-55.6\text{e-}4\pm7.0\text{e-}59.0e-4±1.7e-59.0\text{e-}4\pm1.7\text{e-}52.3e-4±2.9e-52.3\text{e-}4\pm2.9\text{e-}50.323412.5

The conclusion is that exponential integration is a key factor in the accuracy of both methods; however, most of PaPi’s speed-up comes from its parallel implementation. Both methods have comparable accuracies.

We have also added a new ablation "Computational cost and the number of sweeps".

Weakness 3/Question 3:

We believe the remark about strongly nonlinear systems, as well as the question about the size of epsilon, is due to suboptimal presentation on our end. Epsilon was used in the text as a marker of the “scale” of the nonlinear term, to draw analogies with perturbation theory. In the new version of the draft we clarify the discussion and state plainly that PaPi converges unconditionally for all Lipschitz systems, regardless of the strengths of their nonlinearity; it simply converges faster when the Lipschitz constant of said nonlinearity is smaller. We have removed all references to epsilon and state all results in terms only of the nonlinear factor’s Lipschitz constant.

To further emphasize this point, we have included an additional experiment on the inverse problem for the Navier-Stokes equation, a system with notoriously strong nonlinear components. PaPi still achieves state-of-the-art performance with substantially faster training times.

Miscellaneous

We have also used the additional space in the rebuttal to substantially clean up, clarify and illustrate the writing.

I thank the authors for the clarification and additional experiments to illustrate the novelty and effectiveness of PaPi. I am happy to increase the score.

Reviewer_YxW7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PaPi, a parallel-in-time integration scheme for accelerating training and inference in Neuro-Spectral Architectures (NeuSA), which represent time-dependent PDE solutions through spectral coefficients governed by a learned ODE.

The proposed method decomposes the spectral dynamics into an analytically tractable linear component and a learned nonlinear correction. Using Duhamel's formula and Picard iterations, the authors construct updates of the entire temporal trajectory that can be evaluated in parallel through associative scans.

The paper establishes convergence of the continuous Picard–Duhamel iteration under Lipschitz assumptions and provides error estimates for temporal discretization.

Experiments on six forward PDE benchmarks and an inverse parameter-identification problem demonstrate substantial reductions in training time compared with the original sequential NeuSA implementation, generally maintaining competitive or improved accuracy. A comparison against an optimized JAX implementation of RK4 helps isolate the benefits of the proposed integration scheme.

Critical strengths
  1. An effective and mathematically motivated integration strategy.

PaPi addresses a genuine computational bottleneck in neuro-spectral architectures by exploiting the structure of the underlying dynamical system. The combination of exact linear propagation, Picard iterations, and parallel scans enables substantial temporal parallelism while avoiding explicit stability restrictions associated with the linear dynamics. The method is supported by a convergence analysis that clarifies the dependence on the nonlinear correction.

  1. Convincing empirical results and careful implementation comparisons.

The paper demonstrates substantial computational improvements across several PDE benchmarks, frequently with comparable or better accuracy. In particular, the JAX RK4 comparison in Appendix E is valuable because it distinguishes algorithmic improvements from software-framework effects. This substantially strengthens the computational claims.

Critical weaknesses

1. Limited comparison with established exponential integration methods

The principal ingredients of PaPi, including Duhamel's formula, Picard iteration, and exponential propagation, are classical numerical techniques. The main contribution therefore appears to lie in their combination and parallel implementation within NeuSA.

However, the experimental comparisons focus primarily on explicit RK4, whereas established exponential integrators already exploit the same linear/nonlinear decomposition. For example, exponential Euler and exponential Runge–Kutta methods also propagate the linear dynamics exactly, avoiding the stability restrictions of explicit RK4 associated with stiff linear operators.

Consequently, the reported improvements do not fully establish whether PaPi is more efficient than existing exponential integration methods, particularly at matched numerical accuracy. Such a comparison would help distinguish the benefits of exact linear propagation, which is not specific to PaPi, from those of the proposed parallel-in-time approach.

2. Incomplete characterization of numerical and differentiation accuracy

The computational advantage of PaPi depends on the number of Picard sweeps required to resolve the nonlinear correction. This number may increase with nonlinearity and integration horizon, potentially reducing the parallelization benefits.

More importantly, training uses only a few warm-started Picard sweeps, while detaching the trajectory inherited from the previous optimization step. Although the authors establish convergence of the continuous Picard iteration, this does not guarantee that differentiation through a small number of sweeps accurately approximates the derivative of the converged trajectory.

To illustrate the distinction, consider the elementary fixed-point problem

x∗(θ)=θ+0.9x∗(θ),x∗(θ)=10θ.x^*(\theta)=\theta+0.9x^*(\theta),\quad x^*(\theta)=10\theta.

At θ=1\theta=1, initialize the iteration at the exact solution x(0)=10x^{(0)}=10, but detach this value from the computational graph. After two Picard iterations,

x(2)(θ)=1.9θ+8.1.x^{(2)}(\theta)=1.9\theta+8.1.

Thus, at θ=1\theta=1, the forward solution is exact, x(2)(1)=x∗(1)=10x^{(2)}(1)=x^*(1)=10, while the computed derivative is 1.91.9, instead of the exact derivative 1010.

Although this example does not imply that the same discrepancy occurs in PaPi, it illustrates why convergence of the trajectory and accuracy of its derivatives are distinct questions. The paper does not systematically quantify the effect of truncated differentiation on the training gradients.

Finally, Appendix G.2 contains apparent inconsistencies in the exponential quadrature formulas. In particular, setting A=0A=0 in equation (30), the integral on the left becomes

=\frac{t^2}{2}(a+b),$$ whereas the stated right-hand side yields $(a+b)/2$. This suggests missing temporal scaling factors. Also the wording of the sentence "then it the definition of $\varphi$ implies that" seems erroneous. Equation (32) also appears to omit a factor $h$ compared with equation (31). These may be typographical errors, but the authors should clarify the formulas used in the actual implementation.
Questions for authors

Questions for Authors

  1. Comparison with exponential integrators (main concern). Could the authors compare PaPi with exponential Runge–Kutta or exponential time-differencing schemes applied to the same learned spectral dynamics, using comparable accuracy targets and computational budgets? This would better isolate the benefits of time parallelism from those of exact linear propagation.

  2. Accuracy of truncated differentiation (main concern). Training uses only a few warm-started Picard sweeps, with the inherited trajectory detached from the computational graph. How accurately do the resulting gradients approximate those obtained by differentiating the converged trajectory? Could the authors compare the gradients from truncated and converged iterations, particularly on problems with stronger nonlinear corrections?

  3. Convergence and scalability (clarification). How does the number of Picard sweeps needed for a prescribed integration accuracy depend on the strength of the nonlinear dynamics and integration horizon? Could the authors clarify whether integration error contributes to the lower predictive accuracy observed in the 3D layered-wave benchmark?

  4. Exponential quadrature formulas (technical clarification). Equation (30) appears inconsistent when taking A=0A=0, suggesting a missing factor t2t^2; the last occurrence of φ2(A)\varphi_2(A) also appears to require the argument tAtA. Equation (32) seems to omit a factor hh. Could the authors confirm whether these are typographical errors and whether the implemented quadrature uses the corrected expressions?

2 comments

We sincerely thank the reviewer for their comments and insights. We appreciate that they find the theoretical developments and results to be strong; we comment on the weaknesses when answering the questions. Tables have been reduced to save characters. Full versions are in the draft.

Question 1 (Comparison with exponential integrators)

We appreciate this remark. PaPi offers two structural improvements on base NeuSA: using (an analogue of) exponential integration and parallelizable inference, as we emphasize in the draft’s new version. It is fair to compare it with a sequential implementation of exponential integration to control for the effects of each half. In the new draft, we have added an additional ablation to the appendix, comparing PaPi with a sequential 4th-order Exponential Time-Differencing implementation of NeuSA:

ProblemPaPi (rMSE)NeuSA (rMSE)ETD4 (rMSE)PaPi(s)NeuSA (s)ETD4 (s)
3D Layers1.8e-2±1.1e-21.8\text{e-}2\pm1.1\text{e-}27.6e-3±7.7e-47.6\text{e-}3\pm7.7\text{e-}44.8e-3±4.1e-44.8\text{e-}3\pm4.1\text{e-}4251577424
Sine-Gordon5.6e-4±7.0e-55.6\text{e-}4\pm7.0\text{e-}59.0e-4±1.7e-59.0\text{e-}4\pm1.7\text{e-}52.3e-4±2.9e-52.3\text{e-}4\pm2.9\text{e-}50.323412.5

The conclusion is that exponential integration is a key factor in the accuracy of both methods; however, most of PaPi’s speed-up comes from its parallel implementation. There does not seem to be a clear winner in terms of accuracy.

Question 2 (Accuracy of truncated differentiation)

This is a valid concern. In the new draft, we pursue a much more in-depth exposition of the warm starting scheme and position it a key contribution allowing for the use of parallel integration in training.

We address the reviewer’s remarks in two ways. First, we added a full theoretical analysis of the Picard sweeps with warm-start in the form of the new draft’s Theorem 3. In essence, since FF changes only slightly between iterations, the theorem demonstrates that training behaves as expected for a sufficiently large number of warm sweeps, leading to faster convergence of the Picard iteration. Second, we added an ablation experiment to specifically compare the accuracy of the gradients obtained via warm starts. They show that “warm” started iterations achieve similar accuracy to cold ones and their gradients are aligned with the expected full direction as long as sufficiently many sweeps are taken at each step:

ProblemK=1K=1K=2K=2K=4K=4K=8K=8
Sine-Gordon0.999 [0.98]1.000 [1.00]1.000 [1.00]1.000 [1.00]
3D Layers0.308 [-0.35]0.484 [-0.16]0.974 [0.79]--

Caption: Cosine between the gradient through KK sweeps and the exact gradient.

TrainingSine-GordonKdV2D Layers2D Burgers3D Layers
warm, K=2K=2 (default)5.5e-4 ±\pm 8e-51.1e-3 ±\pm 3e-40.070 ±\pm 0.020.029 ±\pm 5e-30.012 ±\pm 3e-3
warm, K=4K=45.6e-4 ±\pm 1e-43.1e-3 ±\pm 8e-40.050 ±\pm 0.030.031 ±\pm 0.026.6e-3 ±\pm 3e-3
cold, K=4K=45.2e-4 ±\pm 1e-40.93 ±\pm 0.040.23 ±\pm 0.050.59 ±\pm 0.010.037 ±\pm 0.01
cold, K=8K=86.5e-4 ±\pm 2e-40.76 ±\pm 0.10.041 ±\pm 0.020.055 ±\pm 0.016.4e-3 ±\pm 2e-3
cold, K=16K=167.2e-4 ±\pm 2e-40.35 ±\pm 0.40.041 ±\pm 0.020.081 ±\pm 0.02OOM

Caption: rMSE with and without the warm start, for KK sweeps per step.

Question 3 (Convergence and scalability)

The first part of this question is partially answered in the draft. By Theorem 1, the error after k sweeps decays like (LT)k/k!(LT)^k/k!, where LL is the Lipschitz constant of the learned term and TT is the time horizon, meaning the number of sweeps needed grows with LTLT. We illustrate this with a new set of ablations. At inference (appendix “Number of Picard sweeps at inference”), eight cold sweeps suffice on every problem, except KdV, whose nonlinear term is largest relative to the linear part, which needs 32 sweeps.

This connects to the answer to the second part. The degradation in accuracy for the 3D may be recovered by increasing the number of sweeps, as seen in the new ablations added to the paper (see above). With four instead of two, the error halves (0.012 to 0.0066) at 1.6 times the training time. Additionally, we show that changing the quadrature rule (trapezoidal or exponential) and the training time grid (full or half) keeps the error between 0.012 and 0019. Integration error therefore contributes, but it is not the main factor. We have corrected the sentence in Section 4 that attributed the gap to the coarse training grid.

Question 4 (Exponential quadrature formulas)

We thank the reviewer for pointing out the inconsistencies in equations (30) and (32). These were typos that we have now corrected in the new draft. The implemented quadrature uses the now correct expression (44).

Dear Reviewer YxW7,

As the discussion period is ending soon, we would like to kindly follow up on our responses above. To address your concerns, we added:

  • An ETD4 comparison to distinguish the benefits of temporal parallelism from those of exact linear propagation.
  • A warm-start convergence analysis and an ablation comparing gradients from truncated and converged iterations, quantifying how additional sweeps improve gradient alignment.
  • Sweep-count, quadrature, and time-grid ablations, showing that additional training sweeps substantially improve accuracy on 3D Layers.

We also corrected the quadrature typos and confirmed that the implementation uses the correct expressions.

We would appreciate hearing whether these responses resolve your concerns and are happy to clarify further in the time remaining. Thank you again for your time and feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.