acceptodds
Under review as a conference paper at ICLR 2027

Predict-Project Transformers Perform Stable Large-Scale Particle Simulation Across Hundreds of Autoregressive Timesteps

Abstract

Simulating free-surface flows with particles is difficult. Particle connectivity changes at every step, and phenomena such as wave propagation emerge only when thousands of particles are coordinated across the whole domain and over hundreds of steps. We find that current learned simulators cannot reproduce these phenomena on core validation scenarios within SPH community. Graph networks are realistic locally but their reach is limited by the receptive range of message passing, while latent-space transformers capture global motion but are bottlenecked by oversmoothing. We also find that no public particle datasets are large enough to clearly measure these failure modes. To address this, we present the Predict-Project Transformer (P2T), which models near-field interactions with message passing and applies a far-field correction with latent attention. On two new datasets built from the SPH community's validation cases, a wave tank and a dam break, P2T is the only model that reproduces the propagating wave and the dam-break impact while keeping particle density close to the reference solver's. Ablations show that its configuration is crucial for rollout stability, and a pre-trained P2T adapts to new tanks, seas and breakers from a handful of simulations. We release the model along with five large-scale datasets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
graph neural network, physics simulation, particle simulation, transformers

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_KxnEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Predict-Project Transformer (P2T), a neural fluid simulator that combines graph neural networks (GNNs) and Transformers in a physics-inspired architecture. Motivated by the predictor-projection decomposition of incompressible fluid simulation, the method uses message passing to capture local particle interactions and latent attention to model global dependencies. Experiments on weakly compressible SPH datasets demonstrate that the proposed method effectively captures both local and global fluid dynamics, outperforming baseline methods on most evaluation metrics.

Critical strengths
  1. The proposed method draws inspiration from the predictor-projection structure of incompressible fluid simulation to guide the integration of local message passing and global attention. Although the corresponding physical properties are not explicitly guaranteed, the experimental results demonstrate the effectiveness of this architectural design.

  2. The ablation experiments demonstrate that the proposed sequential combination of local and global components, together with the skip connection, substantially improves long-horizon prediction accuracy compared with alternative configurations.

Critical weaknesses
  1. The connection between the physical formulation and the proposed neural architecture is not sufficiently validated. The derivation from the physical equations primarily serves as an architectural motivation, while the corresponding physical properties are neither explicitly enforced nor theoretically guaranteed. Specifically, the main connection between Eq. (4) and Eqs. (5)–(7) appears to be the introduction of near-field and far-field features, where the far-field features are conditioned on the near-field features, and the two are combined through a skip connection before decoding the predicted accelerations. However, it remains unclear whether the resulting neural simulator actually approximates the physical decomposition in Eq. (4). For example, does the global attention module learn a physically meaningful pressure-related correction? Do the learned interaction maps or the Jacobian of the complete state-transition function quantitatively resemble those of Eq. (4)? Furthermore, can the method preserve physical properties such as incompressibility, or at least reduce violations of the corresponding constraints? Additional theoretical analysis or quantitative evaluation would strengthen the connection between the physical formulation and the proposed architecture. It would also be useful to clarify whether the architectural design is applicable to other fluid models whose governing equations do not admit the same predictor-projection decomposition.

  2. The physical formulation and the simulation data rely on different fluid models. Eq. (4) is derived from an incompressible SPH (ISPH) projection method, whereas the training datasets are generated using a weakly compressible SPH (WCSPH) solver. Although the authors discuss this distinction and provide a qualitative justification based on acoustic propagation over multiple solver substeps in Appendix C, the equivalence between these two formulations is not fully established. This raises two questions. First, can the proposed method achieve similarly strong performance on datasets generated by an incompressible SPH solver, particularly in terms of maintaining the divergence-free condition? Second, to what extent does the near-field/far-field decomposition derived from Eq. (4) accurately describe the dynamics of the WCSPH datasets? A more quantitative analysis of this correspondence, especially for the more violent dam-break scenario, would help validate the physical motivation.

  3. The mathematical descriptions of the architectural configurations could be clearer. The ablation experiments compare three configurations: P2T, parallel, and no-skip. However, only the proposed P2T configuration is explicitly formulated in Eqs. (5)–(7). Although the parallel and no-skip configurations are described in the final paragraph of Section 4, providing their explicit mathematical formulations would make the architectural differences easier to understand and the experiments easier to reproduce. In particular, it would be helpful to clarify which residual connections are removed in the no-skip configuration, given that the Transformer blocks themselves contain internal residual connections, as shown in Eq. (16).

  4. Some presentation and layout issues reduce readability. Page 8 contains several figures and tables arranged rather densely, making it difficult to identify and follow the key experimental results. A clearer layout would improve readability.

  5. The evaluation of generalization remains limited. First, the transfer-learning experiments focus on related two-dimensional wave-tank scenarios. Although the model is designed to scale to large particle systems, its generalization to more complex three-dimensional fluid dynamics has not been demonstrated. Second, the transfer experiments in Figure 6 primarily report training and validation losses, without corresponding quantitative long-horizon rollout metrics or qualitative comparisons of the predicted dynamics. Such evaluations would provide stronger evidence that the improvements in one-step prediction loss translate into accurate long-term simulation. Finally, videos of the predicted rollouts, particularly for the transfer scenarios, would help assess temporal coherence and the physical plausibility of the simulated dynamics, which are difficult to evaluate from isolated snapshots alone.

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer (1/2)

We thank the reviewer for their thorough feedback on our paper and hope our response can help provide clarity to the reviewer’s concerns.

1. The connection between the physical formulation and the proposed neural architecture is not sufficiently validated.

“The derivation from the physical equations primarily serves as an architectural motivation, while the corresponding physical properties are neither explicitly enforced nor theoretically guaranteed.”

We agree and apologize for the unclear writing. We did not design P2T to learn a specific fluid dynamics implementation, but to use ISPH’s clean structure as a set of inductive biases for a general-purpose fluid network. Our original draft was unclear and we have changed the use of terms such as “confirms” and “is consistent with” to reflect this.

We find empirically that these design choices are supported in P2T. On maintaining near-incompressibility (measured by density deviation), P2T ranks first on wave tank and second on dam break after GNS. However, we note that on both datasets P2T’s ability to maintain near-incompressibility is many times stronger than all other transformer-based baselines, and that configuring P2T’s predictor and projector like in the ISPH equation is crucial for strong overall performance.

Finally, in practice both modern ISPH implementations and WCSPH differ slightly from this analysis even though they all describe the same physical system. In that sense, our approach is similar to how neural operators [1,2] assume that many PDE solutions can be written in terms of integral operators, even though this is not a strict requirement and the actual simulators may do additional computation under the hood.

[1] Li, Zongyi, et al. (2021) Fourier neural operator for parametric partial differential equations. International Conference on Learning Representations

[2] Yang, Zherui, et al. (2026) Simple yet Effective: Low-Rank Spatial Attention for Neural Operators. Forty-third International Conference on Machine Learning

“Does the global attention module learn a physically meaningful pressure-related correction? Do the learned interaction maps or the Jacobian of the complete state-transition function quantitatively resemble those of Eq. (4)? Furthermore, can the method preserve physical properties such as incompressibility, or at least reduce violations of the corresponding constraints?”

Our goal is not to learn a specific operation such as pressure correction but to show that a network that is structurally similar to ISPH can effectively learn fluid simulation generally.

DualSPHysics does not provide gradients, preventing direct computation of reference interaction maps. However, we empirically computed Knear\textbf{K}_{\text{near}} and Kfar\textbf{K}_{\text{far}} from Eq. (4) to determine if P2T’s predictor and projector resemble the structure of Eq. (4). We found that, Knear\textbf{K}_{\text{near}} is banded with nonzero entries only between neighbors of a fixed radius and Kfar\textbf{K}_{\text{far}} is dense. This is trivially true for any message passing subnetwork and latent attention subnetwork respectively. Below we also report the singular values of the interaction map, normalized by the largest singular value. We see that Kfar\textbf{K}_{\text{far}} is effectively rank 2 while Knear\textbf{K}_{\text{near}} has a much higher rank. This aligns with the P2T interaction maps shown in the last page of the appendix, where the predictor contributes a target-specific ring while the projector contributes a global pattern that is shared across all targets.

k12345678
K_far1.00000.90920.05570.00690.00260.00100.00050.0003
K_near1.00000.76840.68880.63730.58560.54020.51660.4963

Finally, as stated above, we find that P2T does achieve a level of near-incompressibility similar to the ground truth simulator and drastically above other attention-based networks. This is reported in Table 2 of the main manuscript.

“It would also be useful to clarify whether the architectural design is applicable to other fluid models whose governing equations do not admit the same predictor-projection decomposition.”

P2T already demonstrates strong performance on WCSPH which does not admit an explicit predictor-projection decomposition. Our experiments show that P2T’s recreated dynamics are by far most physically accurate, and we also find P2T surpasses the existing state of the art on a 3D incompressible formulation of SPH (results discussed below).

AuthorsOpenReview ↗

Response to Reviewer (2/2)

2. The physical formulation and the simulation data rely on different fluid models.

“Can the proposed method achieve similarly strong performance on datasets generated by an incompressible SPH solver, particularly in terms of maintaining the divergence-free condition?”

Recently published work used a 3D dataset generated using divergence free SPH (one type of incompressible SPH method) for training [3]. After training on their published dataset and comparing to their model checkpoint, we find both qualitatively and quantitatively that P2T generates the most accurate rollouts, consistently outperforming the state of the art and dramatically outperforming their baselines. We provide the quantitative comparison below, have included the qualitative visual comparisons in the revised paper’s appendix, and have included video rollouts at https://sites.google.com/view/predict-project-transformers. We are also currently training the baseline models from our paper on this dataset for a more extensive comparison.

Model0.5 s1 s2 s3 s4 s5 s6 s
WorldParticle0.0008860.008550.06170.2260.2700.3000.325
P2T0.0003200.004650.05200.2050.2270.2510.273

[3] Wang, Caoliwen, et al. (2026) WorldParticle: Unified World Simulation of Lagrangian Particle Dynamics via Transformer. To appear in SIGGRAPH ASIA 2026.

“Second, to what extent does the near-field/far-field decomposition derived from Eq. (4) accurately describe the dynamics of the WCSPH datasets?”

Instead of using a predictor-projection decomposition, WCSPH propagates acoustic and material forces locally at dramatically different speeds. These two speeds effectively define local and global interactions when sampled at larger time steps. For example, in one step in the dam break dataset, acoustic information propagates through most of the tank (3.4 m in a 4 m tank) while material propagates 0.035 m.

Since WCSPH and ISPH describe the same fundamental system, we expect that a general architecture capable of learning the dynamics of one solver should also be able to learn the dynamics of the other. We verified this using WCSPH data (wave tank and dam break) and the newly acquired ISPH data mentioned above.

3. The mathematical descriptions of the architectural configurations could be clearer.

We agree that the motivation for the ablations could have been clearer, and we have clarified this in the revision. Equations for each ablation are in the revised paper's Appendix A. Here we provide explanations of the relevance of each ablation.

Both ablations are small routing changes that do not change parameter counts. In the parallel configuration, both the predictor and projector use the encoded features and their outputs are summed for the decoder. This ablation tests if local and global operations alone are enough, or if the predictor outputs provide crucial information to the projector.

In the no-skip configuration, we remove the direct residual connection from local outputs to the decoder. However, there are still per-layer residuals in every block (message passing, encoder, attention) that allow for the uninterrupted propagation of residual activations throughout the entire network. Notably, the no-skip configuration is standard for hierarchical networks such as UPT and GIOROM, and P2T differs from this configuration by one residual connection.

We find that ablating either increases rollout error by >2.8x, meaning that both of the ablated connections are crucial for rollout stability.

4. Some presentation and layout issues reduce readability.

We appreciate this feedback and have made organizational improvements. All figures appear next to the text that references them exception for appendix figures which are explicitly labeled. We also made a small change so that all figures appear in the same order that they are referenced. We hope that this leads to a smoother reading experience.

5. The evaluation of generalization remains limited.

“Although the model is designed to scale to large particle systems, its generalization to more complex three-dimensional fluid dynamics has not been demonstrated.”

As mentioned above, our revised paper includes training results on a 3D SPH dataset where we do much better than the state-of-the-art. Furthermore, we would like to emphasize that our 2D simulations contain more particles than the largest 3D fluid datasets we are aware of, including this new dataset.

“The transfer experiments in Figure 6 primarily report training and validation losses.”

We have added qualitative comparisons that we hope are sufficient to showcase P2T’s ability to learn from a few examples. The paper now includes comparisons of P2T rollouts before and after finetuning. Videos comparing performance before and after fine-tuning can be found at https://sites.google.com/view/predict-project-transformers

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

The paper presented a predict-project transformer (P2T) for fluid simulation. It models near-field interactions with message passing and applies a far-field correction with latent attention.

Critical strengths

The paper is informative and provide a lot of supporting material. The method is sound.

Critical weaknesses

The paper claim the P2T is the only available method reproduces the propagating wave and the dam-break impact while keeping particle density close to the reference solver’s. There are a few good works in these field in the recent years, esp. in 2026, FluidFormer: Transformer with Continuous Convolution for Particle-Based Fluid Simulation Neural Networks, 2026 diffSPH: Differentiable Smoothed Particle Hydrodynamics Journal of Computational Physics, 2026

The paper seems focused on 2D experiments, such as break dam and wave tank, there is little evidence to demonstrate its generalisation and can be extended to other 3D cases. Fig 13 seems have no further implementation details and information to support the outcomes.

Questions for authors

how can the method extended for general 3D simulation? how the work is different from FluidFormer: Transformer with Continuous Convolution for Particle-Based Fluid Simulation

1 comment
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for their feedback, especially for bringing two related papers to our attention. They have been added to the related works.

Extending to 3D simulation

All of our layers work generally, so P2T can be trained on 3D data by providing the third dimension as an input and output feature. 2D datasets were originally our focus since we wanted a close comparison to the related works. For that reason, 3D data was only in our steady-state task (ShapeNet-Car).

We have since added a large 3D dataset that was released just last month by a paper in the related works (WorldParticle paper released May 2026, dataset and code released Oct 2026). Here we found that P2T is both more accurate and faster than the existing state of the art. Below we provide rollout errors compared to WorldParticle's released checkpoint (1 s = 100 rollout steps).

Modelms/step0.5 s1 s2 s3 s4 s5 s6 s
WorldParticle35.10.0008860.008550.06170.2260.2700.3000.325
P2T15.00.0003200.004650.05200.2050.2270.2510.273

Related works

One major advantage of neural simulators is that they can be fine-tuned on real world data to produce more real dynamics than any numerical simulator. Since diffSPH is a numerical algorithm, it cannot learn from real-world data. Furthermore, we expect diffSPH to be slower than P2T since P2T already runs faster than a highly-optimized GPU simulator (DualSPHysics) that is industry-standard.

We thank the authors for pointing us to FluidFormer, as it has a novel approach for propagating local and global information throughout the network. While the method is interesting, FluidFormer's runtime scales quadratically with respect to the number of particles unlike P2T which scales linearly. Therefore, P2T is able to train on much larger datasets than FluidFormer and other particle-level attention networks. FluidFormer has also not released any open-source implementation (the paper's Github link returns a 404 error: github.com/Nianyi-Wang/FluidFormer).

Reviewer_9DqBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces P2T, a particle fluid dynamics model whose architecture is informed by separating the underlying dynamic into local / global interactions, corresponding to a model's near-field and far-field modules. P2T combines near-field message passing with far-field latent attention, enabling accurate prediction of both particle-scale and system-scale phenomena.

Critical strengths

The motivations are clear and model architecture is firmly based in fluid dynamic itself. The experiment result section is well-oriented around observations and conclusions, with direct links to evidence.

Critical weaknesses

I don't have full confidence that P2T's gains stem from architecture alone. As I state in the first two questions, it would be better if the authors provide more detailed comparisons between methods and extend on model ablation on P2T.

Questions for authors

1. Comparability of the reported results

The paper's main claims are comparisons: P2T against graph networks, latent-space transformers, and the reference solver. The paper doesn't report enough to tell how much of each gap comes from the architecture rather than from differences in model size, compute, tuning or evaluation protocol.

I would appreciate it if the authors:

  1. Report parameter counts for UPT, Erwin and GIOROM, and for P2T and GNS on ShapeNet-Car.
  2. Report training cost for every model: wall-clock time, GPU-hours, GPU count and batch size. Put the per-step time and memory at training size from Fig. 14 in a table.

2. Predictor/projector split sweep

Can the authors extend the split sweep (Tables 9–10) toward both ends, for example to 5–10% and 90–95% of parameters in the predictor? I'd like to see the error rise as either component's share goes toward zero, which would show that both are needed. Currently the best split (75%) is at the edge of the sweep.

3. Is there a reason why GIOROM's results were not presented in the main text?

4. Additional literature

There is a recent paper (https://arxiv.org/abs/2511.22938) that uses a GNN+CNN hybrid architecture in place of a traditional multi-scale GNN. I'm interested in the author's positioning of this paper relative to P2T. One specific question is: does this also fall under models that "rely on a set of precomputed hierarchies", thus "leading to poor generalization on new heirarchies"? To be clear, I'm not asking for additional experiments here; the settings are too different and adaptation is too much work.

4 comments
AuthorsOpenReview ↗

Response to Reviewer (1/2)

We thank the reviewer for their thoughtful feedback regarding training details, sweep clarifications, and ablations.

1) Comparability of reported results

We first note that all models are evaluated under exactly the same settings (same test data, same rollout lengths, etc.). Transolver and GNS parameter counts were as close to P2T as possible – differences are only because of rounding up to the nearest multiple of 8 for GPU efficiency (see tables below). For the other baselines, we also went to great lengths to ensure as fair a comparison to baselines as possible, as we detail below.

P2T is on the lower end of parameter counts relative to baseline models. We found its performance was robust across a range of hyperparameters and thus P2T only differs by hidden dimension size across datasets. We trained all models with and without a small amount of training noise and found it was helpful at reducing overfitting for all models. The following lists baseline-specific changes we made for models that performed extremely poorly on our datasets out of the box.

UPT

  • We varied the number of supernodes in [256, 1024, 2048] and found the best performance at 2048.
  • The authors' implementation randomly samples 50–100% of points at each time step. We also tried sampling all points every time (in case this choice was for computational efficiency), but found that this encouraged overfitting.
  • We found that push-forward training did not improve training or rollout loss.
  • Although the authors claim that UPT can perform rollouts in latent space, we found accuracy was significantly higher when decoding and re-encoding at each step.

Erwin

  • The authors provide models with 2- and 4-level hierarchies; we found better performance with the latter.
  • We also tried increasing the hidden width of the 4-level model by 1.5x to create a model with more parameters than all tested P2Ts. However, this model had a worse validation loss, so we did not use it.

GIOROM

  • We found extremely poor performance training with Adamax (used in the original paper), and better performance with AdamW combined with gradient clipping.
  • We swept the particle neighbor radius over [0.03, 0.05, 0.07] and found the best performance at 0.07.
  • Our datasets span a much larger spatial domain. We found better performance fixing the latent grid spacing and scaling the number of nodes than fixing the number of nodes and scaling the spacing.
  • Surprisingly, we observed poor performance training the full-order model (the authors' default on 2D datasets), and found better performance using a randomly-sampled subset of the data (20–25%, following the authors' default on 3D datasets).

While our search was not exhaustive, we emphasize that each of the final models trained for the baseline comparison performed just as well as or (usually) much better than the default code.

Parameter reporting

All models use a batch size of 16 for dam break and 4 for wave tank, and all training was distributed across 4 H100s. The following table has parameter counts and training time for dam break and wave tank training. There was minimal training overhead so GPU-hours is equal to 4 ×\times wall-clock hours.

ModelWave tank ParamsWave tank GPU-hoursDam break ParamsDam break GPU-hours
GNS2.88M22.67.84M31.6
Transolver2.87M6.27.82M8.0
UPT12.6M7.312.6M14.5
Erwin4.75M72.34.76M150.4
GIOROM3.49M11.53.49M12.9
P2T (ours)2.92M25.48.10M33.2

On ShapeNet-Car, GNS and P2T have 3.8M parameters to match the parameter count of the Transolver model published in the authors’ Github repository.

We will also put the data from Fig. 14 in a table for the paper. It is also included here for convenience.

Time per frame (ms) at tank length

Model1×2×4×8×16×
DualSPHysics4685607388461246
GNS23.065.5151.1321.8662.3
Transolver5.713.228.157.9117.4
UPT9.316.230.061.2132.8
Erwin22.546.2139.4290.6OOM
GIOROM14.930.161.3137.8357.1
P2T (ours)23.164.2146.1309.3635.4

Peak GPU memory (GB) at tank length

Model1×2×4×8×16×
DualSPHysics0.0040.0100.0240.0510.105
GNS1.715.0811.925.552.8
Transolver0.210.521.142.384.87
UPT0.290.711.533.186.47
Erwin0.441.214.5617.6OOM
GIOROM0.641.834.188.9120.7
P2T (ours)1.514.4610.422.446.4
AuthorsOpenReview ↗

Response to Reviewer (2/2)

2) Predictor/Projector split sweep We were also interested in understanding the extreme points to determine if both global and local layers were required. However, we weren’t sure about the usefulness of studying extremely underparameterized subnetworks (values like 5% and 95%). We decided that it was more fruitful to compare to Transolver and GNS since they are almost exactly P2T at sweep endpoints (0% and 100% respectively) and more established models.

We made sure to control their hyperparameters for a fair comparison. In the dam break and wave tank experiments the 5 models that make up the sweep (Transolver, GNS, and 3 P2T split models) have parameter counts within 2% and 4% of each other respectively. The variation comes from rounding hidden dimension sizes up to the nearest multiple of 8 for GPU efficiency reasons, and was deemed small enough to allow.

Our results show that P2T at a 75/25 split is a more efficient allocation of parameters than either of its endpoints. For convenience, we have provided a table showing parameter counts as well as rollout errors on wave tank and dam break. Additionally, we qualitatively found that P2T generated more realistic rollouts than either endpoint model (GNS, Transolver) in the paper's Figures 2 and 4.

Wave tank

ModelParamsn=12050100190
Transolver (0%)2.87M0.3459.6242.670.7131.7
P2T (25%)2.89M0.2175.5917.851.7135.0
P2T (50%)2.78M0.2034.9623.040.7121.1
P2T (75%)2.92M0.2024.9215.434.481.3
GNS (100%)2.88M0.2127.5635.0107.5188.9

Dam break

ModelParamsn=12050100190
Transolver (0%)7.82M4.8431899110531033
P2T (25%)8.07M2.962488339811040
P2T (50%)8.13M2.98325826881987
P2T (75%)8.10M2.95260790852964
GNS (100%)7.84M2.472238229891108

3) Why GIOROM is not reported in the main text

GIOROM’s results weren’t in the main text because we had only run it with a single seed at time of submission. However, we’ve since had time to run across multiple seeds. The revised paper will include GIOROM in the baseline comparisons throughout the entire text.

4) Additional literature

We thank the reviewer for bringing this paper up, and it has been added to the related work. Although we have not trained CORGI, we expect its use of a fixed grid hierarchy would lead to poor generalization when the particle-to-grid hierarchy is sufficiently different. We experimentally see this in GIOROM, a similar particle-to-grid model, where past a certain rollout step rollouts either explode or collapse.

Ablations as a critical weakness

The motivation for the ablations could have been clearer, and we have clarified this in the revision. Both ablations are small routing changes that do not change parameter counts. In the parallel configuration, both the predictor and projector use the encoded features and their outputs are summed for the decoder. This ablation tests if local and global operations alone are enough, or if the predictor outputs provide crucial information to the projector. In the no-skip configuration, the operators are composed but there is no direct residual connection from local outputs to the decoder. However, there are still per-layer residuals that allow for an uninterrupted propagation of residuals throughout the entire network. Notably, the no-skip configuration is standard for hierarchical networks such as Erwin and GIOROM, and P2T differs from this configuration by one residual connection. We find that ablating either increases rollout error by >2.8x, meaning that both of the ablated connections are crucial for rollout stability.

We sincerely thank the reviewer for their feedback and time, and are incorporating this information into the revised paper to make our write-up as clear and sound as possible.

Reviewer_9DqBOpenReview ↗

Response to Rebuttal

On the predictor / projector split sweep, what I wanted to see is a finer grid on the percentage of parameters, for example an 8-fold split instead of 4-fold. This would better showcase, in my view, the importance of both components and better support the method's soundness. The ideal result will show a sweet spot for the ratio. I apologize for not being precise enough and for not clarifying sooner.

In general, my assessment for the paper has not changed; I think this paper provides a valuable method informed by the underlying physics, and lists good evidence for the method's superiority within the literature. I maintain my score but lean towards acceptance.

AuthorsOpenReview ↗

Response to reviewer

That is a useful experiment, as split percentage is a hyperparameter we only coarsely swept. Thank you for the response and affirmation of our work's value, we appreciate it.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Review Check-in

We would like to thank the reviewers again for their reviews and ask if they have been adequately addressed, or if not, what further concerns they have.

Dear Reviewers,

Thank you for your reviewing service. The authors have provided detailed rebuttals to the comments. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

Thank you,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.