acceptodds
Under review as a conference paper at ICLR 2027

Vera: Decoupled Simulation for Efficient and Physically-Grounded World Modeling

Abstract

World models have recently gained significant popularity, largely in the form of video-generation models which are able to generate high-quality diverse videos of various scenes. Due to these abilities, many have tried to apply them to the task of generating synthetic data or planning for robotics applications. However, these trials have so far been limited by the lack of physical accuracy in generated videos. Because these video models are trained purely to predict future video frames, they struggle to learn the underlying 3D dynamics and cues required to generate physically consistent and accurate videos. In this work, we propose Vera, a decoupled approach to world modeling. By splitting the task into two stages, simulation and rendering, and carefully designing the simulation representation and architecture, we are able to train a lightweight simulation model which can learn generalizable physics and can be used for robotics planning in real-time. This model outputs 3D trajectories, allowing us to leverage pre-trained models for high-quality video generation through motion conditioning. Across multiple benchmarks, we demonstrate Vera's ability to simulate complex dynamics, learn generalizable physics laws, and generate high-quality, physically-accurate videos.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
World Models, Video Generation, Simulation, Physics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_myD9OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces VeRA, a world modelling method that first learns physical simulation and then use a renderer for RGB rendering. The authors first train a object-centric simulator using geometric and visual features to output accelerations. Then it uses point splatting or video model for synthesizing the final videos. Experiments on physical-parameter generalization, video generations and robot tasks are performed.

Critical strengths
  • Decouping simulation from rendering is reasonable and sound.
  • The evaluations are comprehensive, the authors evaluates unseen physical parameters, object counts, and rollout lengths, etc.
Critical weaknesses
  • The overall pipeline largely follows prior works PhysCtrl, which trains a trajectory transformer first and use a video model for rendering. I didn't see clear difference in terms of high-level idea.
  • The visual results are quite limited, the videos in the supp show noticeable blurrness and artifacts.
  • Figure 9 shows that the diffusion renderer can violate predicted trajectories, but the paper does not clearly distinguish evaluation of simulation and rendering in its physics benchmarks.
Questions for authors

Are the PISA-Bench and WorldBench results computed from predicted trajectories, point-splatted videos, or diffusion-rendered videos?

2 comments

Thank you for taking the time to review our work and for your thoughtful comments and feedback. We have addressed them individually below:

The overall pipeline is similar to PhysCtrl. What are the key differences?

While our work tackles a similar problem to PhysCtrl, there are a few key differences which set our work apart:

  • Multiple Objects: PhysCtrl simulates only a single object, while our proposed technique is designed to handle complex scenes of multiple interacting objects. In Table 8, we show that on the Language-Table set, our model can handle up to 10 objects gracefully even when it has never seen that many during training.
  • Planning: Vera supports agents with per-step actions, enabling planning applications as shown on PushT and Language-Table. PhysCtrl takes in only an initial force prompt and cannot support multi-step planning. Additionally, PhysCtrl only outputs generated video which adds additional complexity for planning tasks.
  • Speed: Due to Vera’s modular architecture, we can perform planning using only the 5.5M parameter simulation model. PhysCtrl requires the full diffusion sampling procedure per inference.

Visual quality is limited

While artifacts and blurriness in the final videos are undesirable and affect the quality of the outputs, they are a result of the quality of the DiffusionAsShader model which we use without modification in this work. Additionally, Table 1b shows that videos from our model outperform prior video models such as Cosmos 3 and Wan 2.2 on video quality metrics (PSNR, LPIPS, FVD).

There is no evaluation to show rendered videos accurately represent the predicted trajectories

Thank you for raising this important point. In order to judge how well generated videos follow the input trajectories, we run CoTracker3 on generated DiffusionAsShader videos and then compare the output tracks with the 2D projections of the input 3D trajectories and calculate the pixel space error. On the synthetic eval set, which covers both our generated data and data taken from Physion, we find the error is 8.3 ± 4.1 pixels (1 std dev), which tends to be a fraction of the overall motion of each trajectory. In addition, we perform a test on the set of friction variation videos: for each video, we use CoTracker3 and DepthAnything3 to estimate the 3D trajectory of the objects in each video. We can then compare the video stopping distance with the input trajectory stopping distance to track whether the generated videos accurately match the input trajectories. The results are shown below. Across all 10 friction levels, the video distance deviates by an average of 2.71 centimeters and tracks the input trajectory closely, showing that the final videos do preserve predicted trajectories.

FrictionTrajectory Stop DistanceVideo Stop DistanceError
0.182.476.7−5.7
0.253.352.5−0.8
0.342.945.4+2.5
0.435.839.2+3.4
0.531.534.3+2.8
0.627.026.9−0.1
0.721.924.0+2.1
0.818.720.6+1.9
0.913.510.2−3.3
1.013.79.2−4.5
Avg Abs Error----2.71

Are PISA-Bench and WorldBench results from trajectories, point-splats, or diffusion videos?

The PISA-Bench and WorldBench results are all computed from DiffusionAsShader videos using the estimation techniques from the original works.

Please let us know if you have any further questions. Thank you again for taking the time to review our work!

Dear Reviewer myD9,

Thank you for your thoughtful review and comments on improving our work. As the discussion period comes to a close, we wanted to kindly follow-up to ask whether our response has addressed your comments and questions. We have responded with additional details on how our work differs from PhysCtrl, details on visual quality limitations and evaluation procedures, and an evaluation of whether our generated videos match the predicted trajectories. If any details of our work remain unclear or if you have any further questions or comments, we would be very grateful for the opportunity to discuss them with you before the end of the discussion period.

Thank you for your time and thoughtful consideration!

Reviewer_V8pCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Vera is a world modeling framework that decouples dynamics simulation from video rendering. It encodes images, depth, segmentation, and physical properties into object-centric 3D tokens. A roughly 5.5M-parameter Transformer predicts accelerations, which are integrated using semi-implicit Euler integration to produce trajectories. These trajectories then condition either a point-cloud renderer or a pretrained diffusion model. The architecture also incorporates FiLM conditioning on physical properties and sigmoid attention with a relative-geometry gate. The paper evaluates trajectory prediction, physical parameter extrapolation, video quality, and robotic planning using synthetic physics scenes, PISA-Bench, WorldBench, PushT, and Language-Table. Its main contribution is this integrated architecture and the reported predictive and planning performance of its compact dynamics model.

Critical strengths
  1. The architecture has practical value and shows fairly consistent trajectory prediction gains. Restricting learned dynamics to compact, explicit 3D object states allows the same prediction module to support planning and different renderers. In Table 1a, Vera achieves the lowest per-step position error in five of six scenarios. In Table 3b, its moving-object MAE and endpoint error are 1.51 cm and 1.73 cm, respectively, outperforming the listed baselines. These results support the usefulness of the representation and prediction approach for the tested rigid-body tasks, although they do not establish independent novelty for every component.
  2. The evaluation covers multiple applications and some out-of-distribution conditions. Beyond video metrics, the paper examines changes in friction and mass, object counts, extended rollout lengths, and robotic planning, alongside architectural ablations. This provides evidence beyond visual demonstrations. In particular, explicit trajectories enable direct inspection of motion errors and responses to physical parameters.
Critical weaknesses
  1. The physical accuracy of the final videos is not established through a sufficiently traceable evaluation procedure. Table 5 does not specify whether gravity is estimated from network predictions, integrated 3D trajectories, or rendered videos. Tables 4 and 5 also omit the renderer configuration and complete measurement procedures. This distinction directly affects the central claim of physically accurate video generation: Figure 9 shows that the diffusion renderer can alter object rigidity and deviate from its conditioning trajectories. The reported video quality metrics do not directly measure this deviation. Consequently, the paper does not adequately establish whether accurate simulated trajectories consistently produce physically accurate videos.
  2. The comparison conditions for planning performance and real-time operation are insufficiently specified. The robotics experiments do not fully describe each method’s observations, goal inputs, physical parameter sources, cost functions, or complete planning budgets. Evaluation counts and the meaning of error bars are also missing. This makes the stability of the PushT difference between 80 and 78 difficult to assess. Language-Table shows larger gains, but the comparison conditions still require clarification. Vera also scores below BC on the two standard tasks, although the training data sizes differ. Finally, the 48-frame simulation timing in Table 9 does not establish real-time operation of a complete planning cycle that includes candidate search and scoring. These omissions limit confidence in both the source of the planning gains and the claimed deployment efficiency.
Questions for authors
  1. Which renderers were used for Tables 4 and 5? For Table 5, which stage’s output was used to estimate gravity, and what scale calibration, fitting procedure, and sample filtering were applied? Please specify the sample count and meaning of “±”. Do you have quantitative measurements of consistency between the rendered motion and the conditioning 3D trajectories?
  2. In PushT and Language-Table, what observations, goals, and physical properties did each method receive? Please specify the cost functions, candidate counts, optimization iterations, and evaluation episode counts, and resolve the inconsistent PushT training set size. On the same hardware, what is the latency of a complete planning cycle, including perception, search, and scoring?
  3. Does the attention ablation in Table 10 retain the geometric gate G? Does the FiLM ablation retain access to the same physical properties? Please identify the evaluation dataset and exact changes in each ablation so that the contributions of the individual components can be distinguished.
4 comments

Thank you for taking the time to review our work and for your thoughtful comments and feedback. We have addressed them individually below:

Which renderers were used for Tables 4 and 5?

DiffusionAsShader was used for both Tables 4 and 5, so the results evaluate the full pipeline including both simulation and rendering. We used the original benchmark's estimation techniques for both Tables.

For Table 5, how was gravity estimated? Please specify sample count and the meaning of “±”

Gravity was estimated using the original benchmarks pipeline, so the calibration, fitting, and filtering were done as in the original paper. Specifically, calibration was done with the checkerboard in the benchmark videos, fitting was done with least squares fitting, and filtering with RANSAC. There were 81 gravity videos in the evaluation set and the ± denotes 1 standard deviation of the estimated coefficients.

What quantitative measurements show final videos preserve trajectories?

Thank you for raising this important point. In order to judge how well generated videos follow the input trajectories, we run CoTracker3 on generated DiffusionAsShader videos and then compare the output tracks with the 2D projections of the input 3D trajectories and calculate the pixel space error. On the synthetic eval set, which covers both our generated data and data taken from Physion, we find the error is 8.3 ± 4.1 pixels (1 std dev), which tends to be a fraction of the overall motion of each trajectory. In addition, we perform a test on the set of friction variation videos: for each video, we use CoTracker3 and DepthAnything3 to estimate the 3D trajectory of the objects in each video. We can then compare the video stopping distance with the input trajectory stopping distance to track whether the generated videos accurately match the input trajectories. The results are shown below. Across all 10 friction levels, the video distance deviates by an average of 2.71 centimeters and tracks the input trajectory closely, showing that the final videos do preserve predicted trajectories.

FrictionTrajectory Stop DistanceVideo Stop DistanceError
0.182.476.7−5.7
0.253.352.5−0.8
0.342.945.4+2.5
0.435.839.2+3.4
0.531.534.3+2.8
0.627.026.9−0.1
0.721.924.0+2.1
0.818.720.6+1.9
0.913.510.2−3.3
1.013.79.2−4.5
Avg Abs Error----2.71

(continued below)

(cont.)

What were the inputs, goals, and hyperparameters of the PushT and Language-Table experiments?

For Language-Table, we use MPPI with a horizon of 12 steps, 200 candidates per iteration, 30 elites chosen with temperature 0.5 and initial variance scale of 1.0. For PushT, we use CEM with a horizon of 25 steps with frame-skip 5, 200 candidates per iteration, 30 elites and initial variance scale of 0.3. We have added these details to the revised PDF as well. We also have put together tables to show the inputs, goal, and cost functions for both scenarios below. For Language-Table, the Image, Depth, and Segmentation inputs are used to create the object-centric inputs and tokens each technique (Ours, IN, NFF, and PF) are trained to expect.

Push-T:

MethodObservationGoalPhysical PropertiesCost Function
OursCurrent state (extracted from image at start) + Agent State + ActionState extracted from ImageN/APosition MSE + 5000 × (1 − cos(Angle Error))
DINO-WM3 Frames + Agent State + ActionImageN/AMSE of Pred and Goal Latent
PLDM3 Frames + Agent State + ActionImageN/AMSE of Pred and Goal Latent
GCBCImageImageN/AN/A
GCIVLImageImageN/AN/A

Language-Table:

ModelObservationGoalPhysical PropertiesCost Function
OursImage + Depth + SegTarget PositionSimulatorMSE between Pred. Position and Goal Pos. + 1e-3 × action magnitude
Ours (vision)Image + Depth + SegText DescriptionSimulatorMSE between Pred. Position and Goal Pos. + 1e-3 × action magnitude
INImage + Depth + SegTarget PositionSimulatorMSE between Pred. Position and Goal Pos. + 1e-3 × action magnitude
NFFImage + Depth + SegTarget PositionSimulatorMSE between Pred. Position and Goal Pos. + 1e-3 × action magnitude
PFImage + Depth + SegTarget PositionSimulatorMSE between Pred. Position and Goal Pos. + 1e-3 × action magnitude
DINO-WMImageTarget ImageN/AMSE between Pred. Latent and Goal Latent + 1e-3 × action magnitude

What is the latency of a complete planning cycle?

We have measured the full latency of a planning cycle including tokenization, candidate generation, inference, and scoring and show the results for both PushT and Language-Table below: Overall, the full cycle takes ~97 ms for Language-Table and 47 ms for PushT, making it fast enough to run in the loop. We have added these tables in the revised PDF to the Appendix as Table 10.

Language-Table:

StepTime (ms)
Tokenize5.2
Candidate Generation0.36
Actions to Trajectories Conversion0.09
Model Inference91
Scoring0.1
Total96.75

PushT:

StepTime (ms)
Tokenize2.1
Candidate Generation0.31
Action to Trajectory Conversion12.8
Model Inference31
Scoring0.9
Total47.11

What were the evaluation counts for PushT and Language-Table? Please resolve inconsistent PushT dataset size.

PushT had 4,000 evaluation samples from 40 unique configurations and Language-Table had 50 evaluation samples per task for 150 total. We apologize for the error on the PushT dataset size. There are 200 unique configurations with 100 noise variants each for 20,000 total samples in the dataset, which we split 80%/20% by configuration.

cont. below

PushT results are missing error bars and meaning of error bars is not clear

We have updated Table 3a in the revised PDF to include error bounds. The error bounds now show 1 standard deviation across 5 unique seeds for both PushT and Language-Table. With error bars, the range of our technique and PDLM overlaps, while our technique remains better than all other baselines.

Does the attention ablation retain the geometric gate G?

In the original table, the gate G is removed alongside the sigmoid attention. To address this gap, we have re-run the ablation study to isolate the sigmoid attention and gate G. We have updated Table 11 in the paper and also show the results below:

MetricMAE (Moving) ↓MAE (All) ↓Endpoint Err. ↓
Ours1.510.901.73
w/o Sigmoid Attn.1.771.232.85
w/o G Gate1.631.412.11
w/o FiLM Encoder1.581.391.93
w/o Acc. Pred2.510.723.71

Does the FiLM ablation retain access to physical properties?

Yes. When we remove FiLM during the ablation, we instead directly concatenate the physics properties into the object token. We have added details to make this clear in section A of the Appendix.

What dataset was used for the ablation?

The Language-Table dataset was used for the ablation study. In order to further identify the individual contributions of each piece of our architecture, we have put together a more comprehensive ablation table below in which we systematically add each component to a base model (object-center tokens, acceleration prediction, and trajectory transformer). For this table, we provide results on the Synthetic Set (both trajectories and videos) and the Language-Table dataset. We have also added this to the revised PDF as Table 12.

ModelSynthetic: MAE (Moving)Synthetic: MAE (All)Synthetic: Endpoint Err.Synthetic: PSNRSynthetic: LPIPSLanguage-Table: MAE (Moving)Language-Table: MAE (All)Language-Table: Endpoint Err.
Base0.14310.13920.505816.240.2042.872.044.76
+FiLM Encoder0.11610.09350.451817.530.1932.091.122.93
+Sigmoid Attn0.10980.10170.462117.170.1981.991.302.77
+Geometric Gate0.11980.12050.485517.620.1902.151.223.04
Full Model0.08980.08490.410818.130.1871.510.901.73

Please let us know if you have any further questions. Thank you again for taking the time to review our work!

Dear Reviewer V8pC,

Thank you for your thoughtful review and comments on improving our work. As the discussion period comes to a close, we wanted to kindly follow-up to ask whether our response has addressed your comments and questions. We have responded with:

  • Details of the inputs and estimation procedures for Tables 4 and 5
  • Details, hyperparameters, and latency measurements for the PushT and Language-Table experiments and tables
  • Evaluation of whether our generated videos match the predicted trajectories
  • More details and additional ablations to further clarify our proposed techniques and results

If any details of our work remain unclear or if you have any further questions or comments, we would be very grateful for the opportunity to discuss them with you before the end of the discussion period.

Thank you for your time and thoughtful consideration!

Reviewer_YG11OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Vera separates world modeling into object-centric 3D dynamics prediction and visual rendering. A compact Transformer uses object observations and physical attributes to predict accelerations, which are integrated into state trajectories; point-based or pretrained generative renderers then produce visual outputs. The system is evaluated on synthetic physical scenarios and robot tasks, with additional studies of physical-parameter variation and component ablations. Its main practical aim is inexpensive simulation for prediction and planning.

Critical strengths
  1. A useful system-level decomposition. Separating dynamics from appearance allows planning to use a small state predictor without repeatedly invoking an expensive video generator. This is a sensible practical approach, and the robot-task results give the architecture a concrete use case.

  2. Useful investigation and disclosure of limitations. The physical-parameter studies and ablations help characterize the dynamics model. Reporting simulation and full-generation costs separately, and showing rendering failures, helps readers understand where the approach is useful and where it breaks down.

Critical weaknesses
  1. The methodological novelty is limited relative to the scope of the claims. The core design integrates object-centric dynamics, physical-attribute conditioning, acceleration prediction/integration, and existing rendering machinery. The sigmoid attention and conditioning choices provide identifiable design decisions, but the ablations do not yet establish a substantial learning contribution beyond this integration. In particular, predicting acceleration and integrating states supplies a kinematic bias; it does not by itself enforce conservation or physically correct contact. A clearer account of what is newly enabled by the learning method, with a suitably matched dynamics alternative, is needed to justify the claimed advance.

  2. The stronger physical-world-model claims exceed the validation. The system receives depth, segmentation, physical attributes, and initial-state information whose availability differs from video-model baselines, making it difficult to separate architectural benefit from information advantage. The estimated-parameter robustness study is useful, but does not fully settle this comparison. Moreover, accurate predicted 3D trajectories do not establish that the final rendered videos preserve those trajectories or their physical properties; the disclosed generative-rendering failures in Fig.9 illustrate precisely this gap. The paper should distinguish state-prediction accuracy, rendered-video fidelity, and planning utility. Until these are evaluated under explicit, comparable input conditions, the broad claims are not sufficiently supported.

Questions for authors
  1. Contribution clarification: Which methodological mechanism provides an advantage beyond a conventional object-centric dynamics predictor with the same inputs, parameter budget, integration scheme, and renderer? What matched evidence isolates that mechanism?
  2. Assessment-changing evaluation: How do results change under comparable input information, including estimated rather than simulator-provided physical attributes? Separately, what quantitative evidence shows that final rendered videos preserve the predicted trajectories and physical relationships, rather than only appearing plausible?
  3. Dataset clarification: Please reconcile the stated 20,000 rollouts with 2,000 starting configurations and 100 variants, and the differing scene-count descriptions. What are the actual training counts and configuration-level train/test splits? This matters for the data-efficiency and generalization claims.
4 comments

Thank you for taking the time to review our work and for your thoughtful comments and feedback. We have addressed them individually below:

Which methodological mechanism provides an advantage beyond a conventional object-centric predictor?

Thank you for the key question. In order to isolate the contributions of each mechanism, we have run another ablation study. For this study, we start with the “Base” model which includes: the object centric tokens, acceleration prediction + semi-explicit Euler integration and DiffusionAsShader. We then add a single mechanism at a time to this base model and train it on the synthetic dataset and Language-Table and evaluate in terms of 3D trajectories on both and generated videos on the synthetic dataset. These results are shown below and show that each mechanism provides unique benefits: the FiLM Encoder greatly helps with static objects while also improving overall, while the sigmoid attention and geometric gate provide improvements on both types of objects. It also shows that the combination of all mechanims leads to the best performance. We have also included this in the revised PDF as Table 12.

ModelSynthetic: MAE (Moving)Synthetic: MAE (All)Synthetic: Endpoint Err.Synthetic: PSNRSynthetic: LPIPSLanguage-Table: MAE (Moving)Language-Table: MAE (All)Language-Table: Endpoint Err.
Base0.14310.13920.505816.240.2042.872.044.76
+FiLM Encoder0.11610.09350.451817.530.1932.091.122.93
+Sigmoid Attn0.10980.10170.462117.170.1981.991.302.77
+Geometric Gate0.11980.12050.485517.620.1902.151.223.04
Full Model0.08980.08490.410818.130.1871.510.901.73

How do results change under comparable input information?

To evaluate the quality of predicted trajectories and videos under comparable input information, we evaluate our model on the synthetic set with VLM-estimated physics parameters and estimated depth and segmentation. Results are shown below and have also been included Table 1b in the revised PDF. We find that even with fully-estimated inputs, generated videos from our model are comparable to Cosmos 3, which takes in 5 frames as input, in PSNR, LPIPS, and FVD, suggesting that our model is robust to in-the-wild scenarios where ideal input is not available.

ModelPSNR ↑LPIPS ↓FVD ↓
Cosmos 317.230.191176.1
Wan 2.214.120.229223.7
Ours (Splat)15.980.208203.6
Ours (Est. Input)17.140.196168.2
Ours18.130.187141.5

What quantitative evidence shows final videos preserve predicted trajectories?

Thank you for raising this important point. In order to judge how well generated videos follow the input trajectories, we run CoTracker3 on generated DiffusionAsShader videos and then compare the output tracks with the 2D projections of the input 3D trajectories and calculate the pixel space error. On the synthetic eval set, which covers both our generated data and data taken from Physion, we find the error is 8.3 ± 4.1 pixels (1 std dev), which tends to be a fraction of the overall 2D motion of each trajectory. In addition, we perform a test on the set of friction variation videos: for each video, we use CoTracker3 and DepthAnything3 to estimate the 3D trajectory of the objects in each video. We can then compare the video stopping distance with the input trajectory stopping distance to track whether the generated videos accurately match the input trajectories. The results are shown below. Across all 10 friction levels, the video distance deviates by an average of 2.71 centimeters and tracks the input trajectory closely, showing that the final videos do preserve predicted trajectories.

FrictionTrajectory Stop DistanceVideo Stop DistanceError
0.182.476.7−5.7
0.253.352.5−0.8
0.342.945.4+2.5
0.435.839.2+3.4
0.531.534.3+2.8
0.627.026.9−0.1
0.721.924.0+2.1
0.818.720.6+1.9
0.913.510.2−3.3
1.013.79.2−4.5
Avg Abs Error----2.71

Discrepancies in rollout counts and missing train/test split information

We apologize for the inconsistencies in the data description: for PushT, there were 200 starting configurations and 100 noise variants to make the 20,000 total rollouts. We have fixed this error in the revised PDF. The data was split 80%/20% at the configuration level to avoid any data leakage. This resulted in 16,000 training rollouts and 4,000 evaluation rollouts. We also have clarified in the revised PDF that for the synthetic set, there are four scenarios since the Table Slide category has data both with and without a "pusher" object which acts as agent.

Please let us know if you have any further questions. Thank you again for taking the time to review our work!

Thank you for the detailed response and the additional experiments. The new ablations clarify what each component contributes, and the clarification of the PushT dataset size and configuration-level split addresses my question about the data. I also appreciate the evaluation with estimated inputs and the quantitative measurements of how closely the rendered videos follow the predicted trajectories. These results strengthen the evidence for the settings tested, though I still have some reservations about how broadly they support the paper’s claims about physical accuracy in rendered videos. I will consider the rebuttal alongside the other reviewers’ comments before deciding whether to raise my score.

Dear Reviewer YG11,

Thank you for your response and consideration! If there are any additional experiments, clarifications, or revisions to the paper that would help address your concerns about how broadly our results support the claims on physical accuracy, we would be grateful to hear them and will do our best to provide them before the discussion period closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.