acceptodds
Under review as a conference paper at ICLR 2027

RATIO: A Relation-Aware Token Interaction Operator for Deformable Solid Dynamics

22 views

Abstract

Deep learning has enabled efficient simulation of deformable solids, yet accurately modeling contact dynamics requires capturing both localized physical interactions and long-range deformation responses. Graph-based methods rely on multi-hop message passing to capture long-range dependencies, incurring substantial computational overhead, whereas methods that compress physical states into a small set of latent tokens may lose the fine-grained information needed to resolve local interactions. Incorporating local physical structure directly into compact global representations provides a natural route toward combining localized interaction modeling with global information exchange. For deformable dynamics, however, this requires capturing and distinguishing heterogeneous intra-body and contact relations. Graph-based simulators naturally represent such relations through connectivity and relation-specific edge attributes, motivating a relation-aware formulation for global modeling. In this paper, we propose RATIO, a Relation-Aware Token Interaction Operator that jointly models nodal states, intra-body mesh relations, and inter-body contact relations through token interactions. RATIO aggregates nodal states and the two types of edge features into separate groups of tokens, applies attention across these groups, and maps the resulting information back to update node and edge states. This design allows intra-body and contact relations to directly participate in global evolution. RATIO can further be coupled with local graph modules for enhanced local interaction modeling. Experiments on deformable solid dynamics with contact demonstrate state-of-the-art accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Deformable Solid Dynamics, Neural Operators, Relation-Aware Tokenization

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_4Fh9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RATIO, a token-based global operator for learned deformable solid simulation with contact. Specifically, nodes, mesh edges, and contact edges are pooled into separate groups of slice tokens, which interact through joint attention and are then redistributed back to nodes and edges. RATIO can be used standalone or fused with a local graph module such as MGN. Results on Cavity Grasping and Tissue Manipulation show the lowest rollout error among ten baselines, with comparable inference cost and parameter count.

Critical strengths

The idea of bringing contact relations directly into global token interaction is simple and meaningful. The baselines are comprehensive, covering local, global, and hybrid methods at comparable parameter counts. The gains are clear: standalone RATIO already beats the strongest baseline MGN-T, and RATIO+MGN reduces its error by 40% on both datasets, with similar inference time. The local-global combination study is also informative.

Critical weaknesses
  • Part of the fused gain seems to come from the fusion framework itself. For example, LSM+MGN (3.89) already beats MGN-T (4.41) on CG. It would be good to state this more clearly. RATIO’s advantage over other global operators within the same framework is convincing on TM but smaller on CG.
  • The evaluation is limited to two small benchmarks with rigid-deformable contact only. Larger meshes, which would test the claimed scalability, and deformable-deformable or self-contact would make the paper stronger.

Minor:

  • When citing MPM in the intro, it’d be more appropriate to also cite the original MPM work by Sulsky et al. [1994] (A particle method for history-dependent materials), rather than only GIMP.
  • In the 2nd paragraph of the intro, the paper attributes far-reaching deformation within one step mainly to large time steps. Another probably bigger factor is material stiffness: stiff materials propagate deformation quickly across the body. This could be mentioned.
Questions for authors
  • Training is notably slower (eg. 112 vs 67 ms per step for RATIO+MGN vs MGN-T). How does total training time compare?
  • Since this is a contact problem, have you checked penetration between the gripper and the body during rollouts, in addition to position RMSE?
4 comments
AuthorsOpenReview ↗

Response 1

We sincerely appreciate your thorough assessment and constructive remarks. Your feedback provides valuable guidance for improving the clarity and rigor of the manuscript. In the following, we respond to each of the points you raised.

Evaluation is limited to two small benchmarks with rigid–deformable contact; larger meshes, deformable–deformable contact, and self-contact would strengthen the paper.

Thank you for this constructive suggestion, which has helped us strengthen the experimental evaluation in our paper. We have added experiments on Deforming Plate (DP) [1] and Rice Grip (RG) [2]. DP features irregular tetrahedral meshes, with 400 recorded states per trajectory corresponding to 399 autoregressive prediction steps. RG uses a particle-based representation of deformable objects, extending the evaluation beyond the mesh-based settings of CG, TM, and DP.

Within the time available for rebuttal, we evaluated representative local, global, and hybrid methods—MGN, Transolver, and MGN-T—alongside RATIO-only and RATIO+MGN on these two datasets. We have also added the dataset descriptions and visualizations to Appendix G. Errors are positional RMSE (×10−3\times 10^{-3}; lower is better).

DatasetMGNTransolverMGN-TRATIO-onlyRATIO+MGN
DP21.3216.3416.0613.9916.75
RG30.4430.8325.4624.9823.28

As shown in the table, standalone RATIO outperforms all three evaluated baselines on both DP and RG, including the hybrid MGN-T model. This extends the favorable comparison observed on CG and TM to the two additional benchmarks. Across these new experiments, the lowest error on each dataset is achieved by a RATIO-based model, either used independently or combined with MGN.

On the other hand, the additional benefit of local–global fusion varies between datasets. Adding MGN reduces the error from 24.98 to 23.28 on RG, but increases it from 13.99 to 16.75 on DP. This illustrates that introducing a local branch does not automatically improve an already effective global operator. As discussed in Section 5 (Conclusion and Limitations), a fuller understanding of when and why local and global operators complement each other is still needed to guide fusion designs that yield more consistent improvements.

These results support the central contribution of our work: a relation-aware global operator that delivers strong prediction accuracy on its own. The performance of standalone RATIO on both new benchmarks broadens the evidence for the operator beyond CG and TM. Fusion with a local module can help the model capture additional local information, but the resulting performance benefits vary across datasets.

AuthorsOpenReview ↗

Response 2

Part of the gain appears to come from the fusion framework itself; RATIO’s advantage over other global operators is more convincing on TM than on CG.

Thank you for this thoughtful observation. We agree that the additional gains from fusion should be distinguished from the contribution of the RATIO operator itself.

The main contribution of this paper is a relation-aware global operator that incorporates mesh connectivity and potential point-pair contact directly into global computation. RATIO compresses node states, mesh relations, and contact relations into separate token sets, jointly updates them through global attention, and redistributes the resulting information to node and edge states. Local structural and contact information therefore participates in global interactions within RATIO itself, without requiring an additional local message-passing module. Standalone RATIO already outperforms all evaluated baselines in the revised main comparison on CG and TM (Section 4.2, Table 1), as well as all three baselines on DP and RG (Appendix G, Table 8), demonstrating the effectiveness of the operator on its own.

As described in Section 3.4 (Figure 4), the optional fusion framework couples RATIO with a local graph operator. This explores a design dimension orthogonal to RATIO’s internal relation-aware structure, allowing us to investigate how additional local propagation can further improve RATIO-based models. We regard this coupling as a useful practical strategy and agree that fusion itself can contribute substantial gains.

The fusion results in Section 4.2 (Table 2) illustrate both the potential of local–global coupling and its dependence on the operator pairing and dataset. These original single-run results are distinct from the three-run means now reported in Table 1. Within the same fusion framework using MGN, RATIO achieves errors of 2.51 on CG and 5.45 on TM, compared with 3.89 and 9.07 for LSM, respectively. Adding MGN improves standalone RATIO on CG, TM, and RG, but increases its error from 13.99 to 16.75 on DP. Adding GATv2 also increases RATIO’s error, from 4.10 to 8.37 on CG and from 7.31 to 8.39 on TM. RATIO’s advantage over other global operators is more consistent on TM, while LSM performs better with GATv2 on CG. As discussed in Section 5 (Conclusion and Limitations), when and why local and global operators complement each other remains insufficiently understood, and these mixed outcomes motivate further investigation of operator compatibility and coupling design.

Overall, the central contribution of this paper is RATIO, whose standalone performance already surpasses the evaluated baselines. Fusion offers a promising direction for extending its capabilities, while a systematic explanation of when and why such coupling is effective remains a topic for future research beyond the main focus of this paper.

AuthorsOpenReview ↗

Response 3

The Introduction should also cite the original MPM work by Sulsky et al. (1994), rather than only GIMP.

Thank you for suggesting this valuable reference. We have added Sulsky et al. (1994) to the MPM discussion in Section 1 of the revised manuscript.

The explanation of far-reaching deformation within one step should also acknowledge material stiffness, rather than mainly attributing it to large time steps.

Thank you for this helpful comment. We agree that material stiffness plays an important role in how rapidly mechanical disturbances propagate through a body. Our reference to “relatively large time intervals” was intended as a comparison with the small integration steps required by explicit solvers. For example, a commonly used estimate of the critical time step in explicit finite element analysis is [3]

\qquad c_d=\sqrt{\frac{E(1-\nu)}{\rho(1+\nu)(1-2\nu)}},$$ where $\ell_e$ is the characteristic element length and $c_{d,e}$ is the dilatational wave speed in element $e$. The wave-speed expression above applies to an isotropic, linear-elastic material, with Young’s modulus $E$, Poisson’s ratio $\nu$, and density $\rho$. At fixed density and Poisson’s ratio, **higher stiffness increases the wave speed and reduces the stable explicit time step** [3]. We have revised Section 1 to clarify the role of material stiffness, density, and mesh resolution in the stable time-step limits of explicit solvers. > Training is notably slower per step, and the comparison should also report total training time. Thank you for highlighting this computational trade-off and asking about the total training cost. These timings were measured on a single NVIDIA GeForce RTX 3090 GPU with 24 GB of memory. The reported per-step training time is obtained by dividing the total training time by the number of optimizer updates. As described in Appendix C, we train on CG for 400 epochs with a batch size of 13. With 600 training trajectories and 104 transitions per trajectory (Appendix A), this schedule comprises 1.92 million optimizer updates. The training times reported in Appendix D (Table 6 in the revised manuscript)—111.96 ms for RATIO+MGN and 67.31 ms for MGN-T—therefore correspond to approximately **59.7 hours and 35.9 hours**, respectively. We acknowledge that RATIO’s current implementation incurs higher training costs than several baselines. Processing irregular graph relations introduces a **memory–runtime trade-off**, particularly when graphs in the same batch have different structures and edge counts. In RATIO, world-edge counts vary across samples and frames. A direct broadcast-based aggregation materializes a large edge–token–channel tensor, whereas per-sample matrix multiplication avoids this intermediate at the cost of multiple smaller operations. During training, backpropagation through aggregation and redistribution adds computation and intermediate-storage requirements. We view this as a **broader implementation challenge in mapping variable-sized graphs into latent tokens for Transformer processing**. Even when global attention operates on a fixed token budget, aggregation into tokens and redistribution to graph entities must still accommodate irregular inputs. Possible optimization routes include **masked padding and grouped matrix multiplication**. TokenGT [4] uses padding and masks to batch variable-length node–edge sequences; applying the same batching principle to edge aggregation permits regular batched matrix multiplication, at the cost of storing and processing padded entries. Grouped matrix multiplication (grouped GEMM) [5] offers another route by jointly scheduling matrix multiplications of different sizes without padding every sample to the largest edge count. The relative benefits of these choices may depend on the dataset, particularly the distribution of graph sizes and edge counts, as well as batch size and hardware. We have added a discussion of this overhead and possible implementation optimizations to Appendix D, following Table 6. A systematic evaluation of these optimizations remains future work.
AuthorsOpenReview ↗

Response 4

Gripper–body penetration during predicted rollouts should be evaluated alongside position RMSE.

Thank you for highlighting this contact-related aspect of the evaluation. We evaluated gripper–body penetration on the CG dataset using all 120 test trajectories, and applied the same geometric measure to the ground-truth trajectories.

We define penetration as a deformable node lying inside the solid material of a gripper. Its penetration depth is the shortest Euclidean distance to the gripper's material boundary, and is zero for nodes outside the material. The calculation uses the original gripper surface geometry, preserving its cavities and excluding internal interfaces between overlapping components. For each trajectory, we take the maximum depth over all deformable nodes and predicted states. We report the mean and maximum of these per-trajectory peak depths across the 120 test trajectories. All values are in the dataset's original coordinate units.

MetricGTMGNTransolverMGN-TRATIORATIO+MGN
Mean trajectory peak depth0.0029380.0027860.0041880.0020570.0037760.003221
Maximum depth0.0089300.0082210.0101000.0124810.0141420.012530

First, the CG ground truth itself exhibits penetration under this geometric measure, making it an essential reference for interpreting the predictions. MGN is closest to GT in both reported summary statistics. MGN-T has the lowest mean trajectory peak depth, below the GT value, although its maximum depth remains higher than GT. RATIO and Transolver exhibit higher mean peak depths than GT. Compared with standalone RATIO, RATIO+MGN reduces both the mean peak depth and the maximum depth, bringing both statistics closer to the GT reference.

These observations are consistent with a possible benefit of local message passing for refining contact behavior. MGN directly propagates information along neighboring mesh and contact edges, which may help resolve local gripper–body interactions. RATIO incorporates these relations into global token interactions to capture deformation across the body, and the CG qualitative comparisons in Figure 5 (Section 4.2) and Figure 8 (Appendix E) illustrate improved overall shape reconstruction by RATIO-based models.

We have added the metric definition, quantitative results, and discussion to Appendix F (Table 7) of the revised manuscript.

Thank you again for your thoughtful feedback. Should you have any further questions or concerns, we would be pleased to provide clarification and continue the discussion.

[1] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning Mesh-Based Simulation with Graph Networks. In International Conference on Learning Representations, 2021.

[2] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B. Tenenbaum, and Antonio Torralba. Learning Particle Dynamics for Manipulating Rigid Bodies, Deformable Objects, and Fluids. In International Conference on Learning Representations, 2019.

[3] Harm Askes, Antonio Rodríguez-Ferran, and Jack Hetherington. The effects of element shape on the critical time step in explicit time integrators for elasto-dynamics. International Journal for Numerical Methods in Engineering, 101(11):809–824, 2015.

[4] Jinwoo Kim, Tien Dat Nguyen, Seonwoo Min, Sungjun Cho, Moontae Lee, Honglak Lee, and Seunghoon Hong. Pure Transformers are Powerful Graph Learners. In Advances in Neural Information Processing Systems, 2022. Official batching implementation.

[5] Babak Hejazi. Introducing Grouped GEMM APIs in cuBLAS and More Performance Updates. NVIDIA Technical Blog, 2024.

Reviewer_W2TMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RATIO, a neural model for predicting the dynamics of deformable solids in contact. The main idea is to build separate token groups for nodes, mesh edges, and inter-body contact edges, let them interact through attention, and map the resulting updates back to the graph. RATIO can be used on its own or combined with a local graph network through a learned gate. The authors evaluate both configurations on cavity grasping and tissue manipulation using 104-step autoregressive rollouts. The combination of RATIO and MeshGraphNets achieves the lowest positional errors among the tested methods. Further experiments examine how the choice of local operator, relation types, and token allocation affects performance.

Critical strengths
  1. The full model achieves substantial improvements on the evaluated tasks. In Table 1, RATIO+MGN reduces positional RMSE from 4.41 to 2.51 on CG and from 8.60 to 5.45 on TM relative to MGN-T, corresponding to reductions of 43.1% and 36.6%. Table 4 reports fewer parameters for RATIO+MGN than for MGN-T (2.195M versus 2.858M). These results are evaluated over 104-step autoregressive rollouts, with checkpoints selected by validation rollout error. This provides meaningful evidence of improved trajectory prediction under the reported settings.

  2. The modular comparisons provide useful empirical information. Table 2 evaluates RATIO with three different local operators and reports the lowest error for RATIO in five of the six local-operator/dataset comparisons across the three global operators tested. Table 3 also compares token allocations and removes individual relation types while keeping the total token budget fixed. These experiments support the usefulness of the proposed representation and help identify the dependence of fusion benefits on the local operator.

Critical weaknesses
  1. The central methodological advantage is not sufficiently isolated. As acknowledged in Section 2, incorporating edge information into compressed tokens already has precedent in Unisoma. The distinctive contribution here is the separation of node, mesh-edge, and world-edge token banks followed by joint interaction. Table 3 is relevant, but the Shared variant is insufficiently specified: it is unclear whether it preserves the endpoint-conditioned representations and type-specific projections of Eq. (8), and how it accounts for differences in entity counts. Shared improves from 14.17 to 9.67 when mesh relations are removed, leaving pooling imbalance as a plausible explanation for its poor performance. Moreover, these ablations concern only standalone RATIO on TM, whereas the strongest results use RATIO+MGN. A matched, count-balanced shared-pooling control is needed to establish the benefit of separate relation banks over a simpler alternative. This unresolved attribution limits the methodological contribution.

  2. The physical conditioning of the prediction task is unclear. Appendix A states that elastic properties vary across samples in both benchmarks, but Eq. (3) does not explicitly include material parameters, and the nine point-type categories are not defined. It is therefore unclear whether material properties are encoded in these labels, supplied elsewhere, or left unobserved. This distinction matters particularly at initialization, where the preceding displacement is set to zero. The manuscript also needs to specify whether all baselines receive equivalent material information. Without this clarification, the precise prediction task and the interpretation of the reported comparisons remain uncertain. This is a potentially fixable issue, but it concerns the experimental setup rather than an incidental implementation detail.

Questions for authors
  1. Clarification of the task: What do the nine point-type categories in Eq. (3) represent, and how are the varying elastic properties described in Appendix A handled? Are different materials pooled during training, and do RATIO and all baselines receive the same material information?

  2. Clarification of the main ablation: How exactly is Shared constructed? Does it retain the endpoint-conditioned features and type-specific projections in Eq. (8), and how are the contributions of nodes, mesh edges, and world edges balanced? Are the backbone and token schedule otherwise identical to those of the separate-bank variants?

  3. Evidence that could change my assessment: Does the advantage of separate token banks persist against a shared-pooling control that retains the same endpoint representations, explicitly balances contributions across entity types, and uses the same backbone and total token budget? A comparison for RATIO+MGN on CG would be particularly informative, since the current tokenization ablations do not cover this configuration. Please report variability across independent training runs for this comparison.

5 comments
AuthorsOpenReview ↗

Response 1

We sincerely thank you for your detailed feedback and the effort devoted to evaluating our paper. Your comments are helpful in identifying where further clarification and evidence are needed. We provide our responses below.

The Shared design needs clarification, and stronger matched comparisons are needed to establish the benefit of separate token banks.

Thank you for giving us the opportunity to further clarify the distinction between separate and shared token banks. The Shared variant retains the endpoint-conditioned edge representations and type-specific input projections in Section 3.3 (Eq. (8)), mapping node, mesh-edge, and world-edge features into a common feature space. These projected features are then pooled into a single token bank through a shared slice-assignment map, allowing every entity to contribute to any shared slice. After attention over the shared tokens, the same assignment weights redistribute the updates to the original entities, followed by type-specific output projections for nodes, mesh edges, and world edges. The backbone and layer-wise total token budget are unchanged. No entity-count balancing is applied: all retained entities participate directly in pooling, with normalization by each slice's total assignment weight but no additional reweighting by entity type. We have clarified this Shared design in Section 4.3 of the revised manuscript.

  • New results on CG.

    To further investigate the behavior of Shared, we have extended the ablation study to CG, complementing the TM results in the original submission. These results have been incorporated into Section 4.3 (Table 3) of the revised manuscript and are reproduced below. Errors are positional RMSE (×10−3\times 10^{-3}; lower is better).

    CGTM
    Relations retained6:1:12:1:1Shared6:1:12:1:1Shared
    Nodes + mesh + world4.105.285.957.318.2814.17
    Nodes + world5.218.219.737.748.439.67
    Nodes + mesh15.7415.4816.278.309.1611.80

    The results support two main observations. First, both separate-bank allocations outperform Shared across all three relation settings on both CG and TM, supporting the value of preserving distinct node and relation representations during compression. Second, world edges play a more important role than mesh edges in these ablations: retaining nodes and world edges consistently yields lower errors than retaining nodes and mesh edges across all three allocations on both datasets. For example, under the default 6:1:1 allocation on CG, removing world edges increases the error from 4.10 to 15.74, compared with 5.21 after removing mesh edges. These results highlight the importance of explicitly incorporating potential contact relations into global computation.

AuthorsOpenReview ↗

Response 2

  • Analysis of the TM mesh-removal result.

    The effect of removing mesh relations differs between the two datasets: Shared's error increases from 5.95 to 9.73 on CG, whereas it decreases from 14.17 to 9.67 on TM. The improvement observed after mesh removal on TM therefore does not extend to CG. To further examine this observation, we conducted two additional training runs for each of the two TM Shared configurations within the time available for rebuttal, giving three runs per configuration. The table below reports these additional Shared experiments alongside RATIO-only, with positional RMSE (×10−3\times 10^{-3}) expressed as mean ±\pm standard deviation across three training runs. The RATIO-only result is taken from the revised main comparison (Section 4.2, Table 1), which now includes variability results for all main CG and TM methods. The two Shared configurations are reported here as additional rebuttal experiments.

    ModelRelations retainedTM error
    RATIO-only (6:1:1)Nodes + mesh + world7.06 ± 0.55
    SharedNodes + mesh + world12.98 ± 1.81
    SharedNodes + world9.40 ± 1.55

    The repeated experiments preserve the main trends: RATIO-only outperforms both Shared configurations, while removing mesh edges improves Shared on TM. Our repeated training experiments on TM also show greater run-to-run variability for Shared than for RATIO-only and the main baselines. One possible explanation is that pooling node and edge entities into a single token bank requires the model to implicitly balance their contributions through its learned feature projections and slice-assignment weights. Learning this balance alongside the dynamics may make optimization more difficult and contribute to the larger variability.

    We also observe that removing mesh edges is more detrimental on CG than on TM. This may be related to differences in mesh-edge preprocessing. The supplied dataset edges were not directly extracted from the original meshes. We reconstructed mesh connectivity for CG, but retained the supplied connectivity for TM because of its irregular node distribution. This preprocessing difference may affect the value of mesh relations in shared pooling, potentially explaining why Shared performs worse on TM when mesh edges are included than when only node and world-edge features are used.

  • Effect of entity weighting in Shared.

    To further assess shared tokenization, we added Shared+, an ablation with inverse-count weighting during pooling, described in Section 4.3 of the revised manuscript (Eq. (13)). For each sample, entities of type rr receive the weight

    \qquad (c_{\mathrm{node}},c_{\mathrm{mesh}},c_{\mathrm{world}}) =(6,1,1),$$ where $N_r$ is the number of entities of that type; absent types receive zero weight. Before slice normalization, this gives an aggregate node, mesh-edge, and world-edge weight ratio of **6:1:1**, following the allocation that performed best with all relations retained. These weights are applied to both the pooling numerator and denominator. Shared+ retains Shared's type-specific input and output projections, backbone, token budget, and redistribution. Following your suggestion, we compare Shared+ and RATIO on CG, both with and without the MGN local branch. The results have been added to Section 4.3 of the revised manuscript (Table 4) and are reproduced below as mean $\pm$ standard deviation across training runs. All three entity types are retained; errors are positional RMSE ($\times 10^{-3}$; lower is better). | Dataset | Shared+ | Shared+ + MGN | RATIO | RATIO+MGN | |:---:|:---:|:---:|:---:|:---:| | CG | 5.22 ± 0.43 | 3.35 ± 0.11 | 3.56 ± 0.47 | 2.69 ± 0.16 | These results suggest that carefully adjusting the type weights can improve shared pooling to some extent, but a performance gap remains relative to separate token banks. This gap persists when both operators are combined with MGN, supporting the benefit of separate token banks in both standalone and fused configurations on CG.

Overall, our ablations on CG and TM support the benefit of separate token banks and highlight the importance of world edges.

AuthorsOpenReview ↗

Response 3

The treatment of material information, the nine point-type categories, and input fairness across methods need clarification.

Thank you for raising this point. Our autoregressive simulation setup follows the MeshGraphNets (MGN) framework [1]. The RATIO inputs are specified in Section 3.2 (Eq. (3)) and describe geometry, node roles, motion history, and prescribed boundary motion. In our CG and TM experiments, material parameters such as Young's modulus and Poisson's ratio are not supplied to the models, and no material-identity labels are provided. Trajectories with different elastic properties are pooled within each dataset, with a single model trained per method and dataset.

The nine-dimensional one-hot vector follows MGN's node-type encoding convention [1]. In CG and TM, NORMAL identifies deformable nodes whose motion is predicted, while OBSTACLE identifies gripper or control nodes with prescribed motion. The additional DP experiments also use HANDLE for constrained boundary nodes. These labels describe physical roles and boundary conditions; they do not encode material properties. The dimension of nine is inherited from the general encoding scheme and does not mean that nine categories are active in each benchmark.

For a fair comparison, RATIO and all evaluated baselines use consistent node-role definitions and the same material-information setting. All methods are trained on the same dataset splits, and none receives additional material parameters or material-identity labels. This setting is consistent with established CG/TM baselines, including the standard MGN baseline evaluated by Linkerhägner et al. [2] and the released CG/TM implementations of UniSoma [3]. More directly, the CG/TM benchmark study [2] reports results for MGN without material inputs. Our reported comparisons likewise evaluate prediction with unobserved material properties.

We agree that this leaves uncertainty at initialization, when the preceding displacement is zero and no deformation history is available to help distinguish material responses. We have clarified the material-input setting, training across materials, and node-type definitions in Appendix C (Input Features) of the revised manuscript.

Thank you once again for your careful assessment and constructive comments. We would be glad to address any remaining questions and engage in further discussion.

[1] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning Mesh-Based Simulation with Graph Networks. In International Conference on Learning Representations, 2021. Official node-type encoding.

[2] Jonas Linkerhägner, Niklas Freymuth, Paul Maria Scheikl, Franziska Mathis-Ullrich, and Gerhard Neumann. Grounding Graph Network Simulators using Physical Sensor Observations. arXiv preprint arXiv:2302.11864, 2023.

[3] Shilong Tao, Zhe Feng, Haonan Sun, Zhanxing Zhu, and Yunhuai Liu. Unisoma: A Unified Transformer-based Solver for Multi-Solid Systems. arXiv preprint arXiv:2506.06021, 2025. Released CG and TM input implementations.

Reviewer_oW1KOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RATIO, a relation-aware token interaction operator for simulating deformable solids with contact. To address this problem, RATIO represents node states, intra-body mesh relations, and inter-body contact relations using separate groups of latent tokens.

Critical strengths
  1. improvements are quite convincing on the evaluated benchmarks。
  2. the paper is well-written and easy to follow.
Critical weaknesses
  1. need a clearer explanation of the novelty relative to existing graph tokenization methods.
  2. evaluation is still relatively limited in terms of physical scenarios and scale.
  3. The computational trade-off deserves more discussion. I appreciate that the authors provide runtime measurements. However, RATIO introduces noticeably higher training costs than several baselines.
Questions for authors

Could the authors clarify the main methodological difference between RATIO and existing node-edge tokenization approaches, especially those that already incorporate relation information into global attention?

6 comments
AuthorsOpenReview ↗

Response 1

Thank you for your careful review and helpful suggestions. We appreciate the opportunity to clarify our work and discuss your concerns in greater detail. Below, we address your comments point by point.

The evaluation remains limited in its coverage of physical scenarios and scale.

We appreciate your suggestion to broaden the evaluation, which has helped us improve the empirical support for our paper. We have added experiments on Deforming Plate (DP) [1] and Rice Grip (RG) [2]. DP features irregular tetrahedral meshes, with 400 recorded states per trajectory corresponding to 399 autoregressive prediction steps. RG uses a particle-based representation of deformable objects, extending the evaluation beyond the mesh-based settings of CG, TM, and DP.

Within the time available for rebuttal, we evaluated representative local, global, and hybrid methods—MGN, Transolver, and MGN-T—alongside RATIO-only and RATIO+MGN on these two datasets. We have also added the dataset descriptions and visualizations to Appendix G. Errors are positional RMSE (×10−3\times 10^{-3}; lower is better).

DatasetMGNTransolverMGN-TRATIO-onlyRATIO+MGN
DP21.3216.3416.0613.9916.75
RG30.4430.8325.4624.9823.28

As shown in the table, standalone RATIO outperforms all three evaluated baselines on both DP and RG, including the hybrid MGN-T model. This extends the favorable comparison observed on CG and TM to the two additional benchmarks. Across these new experiments, the lowest error on each dataset is achieved by a RATIO-based model, either used independently or combined with MGN.

On the other hand, the additional benefit of local–global fusion varies between datasets. Adding MGN reduces the error from 24.98 to 23.28 on RG, but increases it from 13.99 to 16.75 on DP. This illustrates that introducing a local branch does not automatically improve an already effective global operator. As discussed in Section 5 (Conclusion and Limitations), a fuller understanding of when and why local and global operators complement each other is still needed to guide fusion designs that yield more consistent improvements.

These results support the central contribution of our work: a relation-aware global operator that delivers strong prediction accuracy on its own. The performance of standalone RATIO on both new benchmarks broadens the evidence for the operator beyond CG and TM. Fusion with a local module can help the model capture additional local information, but the resulting performance benefits vary across datasets.

AuthorsOpenReview ↗

Response 2

RATIO’s methodological novelty relative to existing node–edge tokenization methods, particularly those incorporating relation information into global attention, needs clearer explanation.

Thank you for raising this point. We appreciate the opportunity to clarify how RATIO differs from existing approaches and would like to explain these differences from the following two perspectives.

  • Comparison with existing node–edge neural operators.

    UniSoma [4] uses within-body edge features to enrich the local structural information in body-level slices, pooling node and edge features together into object representations on which attention operates. As described in Section 3.3, RATIO instead represents mesh and contact relations with dedicated tokens that participate directly in global attention and receive global updates. This allows localized contact information to shape the global deformation response, while global context refines both internal and contact-relation representations.

    Whereas UniSoma models contact through body-level representations, RATIO explicitly represents potential contact between points on different bodies using world edges, as defined in Section 3.1 (Eq. (1)). These world edges provide particularly valuable information about the evolving contact structure, with their connectivity and counts changing across frames as bodies move and interact. Through their dedicated tokens, these changing contact relations participate in global computation and receive updates informed by the entire system. This richer representation also brings a computational trade-off, as variable-size edge sets disrupt the regular structure of batched computation and require graph-like aggregation and redistribution, introducing additional costs when training with multiple samples per batch.

    To further assess the contribution of world edges, we have extended the relation-removal ablation study to CG, complementing the TM results in the original submission. Section 4.3 (Table 3) of the revised manuscript now reports both datasets under a fixed total token budget. We reproduce these ablation results below; errors are positional RMSE (×10−3\times 10^{-3}; lower is better).

    CGTM
    Relations retained6:1:12:1:1Shared6:1:12:1:1Shared
    Nodes + mesh + world4.105.285.957.318.2814.17
    Nodes + world5.218.219.737.748.439.67
    Nodes + mesh15.7415.4816.278.309.1611.80

    For the default 6:1:1 configuration, removing world edges increases the error from 4.10 to 15.74 on CG and from 7.31 to 8.30 on TM. The degradation on both datasets supports the value of explicitly incorporating point-pair contact geometry into global computation, with the particularly large increase on CG highlighting the importance of these relations in that setting. In the revised main comparison (Table 1), standalone RATIO also outperforms UniSoma in mean RMSE on both CG (3.56 vs. 17.20) and TM (7.06 vs. 13.84).

  • Comparison with general node–edge attention.

    TokenGT [3] treats individual nodes and edges as tokens and applies a standard Transformer to their joint sequence. RATIO provides a practical implementation of batched latent-token interaction for graphs with different connectivity patterns, compressing features into a fixed number of tokens before global attention.

    More importantly, our empirical results favor separate compression of node states and distinct physical relations in the physics tasks studied here. With all relations retained and the total token budget held fixed, both separate-bank allocations outperform the evaluated Shared variant. The 6:1:1 configuration reduces the error from 5.95 to 4.10 on CG and from 14.17 to 7.31 on TM, while the 2:1:1 configuration also achieves lower errors of 5.28 and 8.28, respectively. These results support preserving distinct node, mesh, and contact representations during compression, while allowing all three to exchange information through joint global attention.

Overall, RATIO differs from existing node–edge neural operators such as UniSoma by explicitly incorporating point-pair contact relations into global attention. Compared with general node–edge attention exemplified by TokenGT, RATIO performs global interaction through compact latent tokens, with separate representations for nodes, mesh relations, and contact relations. Our results on CG and TM support the value of explicitly representing world edges and preserving these distinctions during compression.

AuthorsOpenReview ↗

Response 3

RATIO’s higher training cost relative to several baselines warrants further discussion of the computational trade-off.

Thank you for highlighting this computational trade-off. As reported in Appendix D (Table 6 in the revised manuscript), RATIO’s current implementation incurs higher training costs than several baselines. Processing irregular graph relations introduces a memory–runtime trade-off, particularly when graphs in the same batch have different structures and edge counts. In RATIO, world-edge counts vary across samples and frames. A direct broadcast-based aggregation materializes a large edge–token–channel tensor, whereas per-sample matrix multiplication avoids this intermediate at the cost of multiple smaller operations. During training, backpropagation through aggregation and redistribution adds computation and intermediate-storage requirements.

We view this as a broader implementation challenge in mapping variable-sized graphs into latent tokens for Transformer processing. Even when global attention operates on a fixed token budget, aggregation into tokens and redistribution to graph entities must still accommodate irregular inputs. Possible optimization routes include masked padding and grouped matrix multiplication. TokenGT [3] uses padding and masks to batch variable-length node–edge sequences; applying the same batching principle to edge aggregation permits regular batched matrix multiplication, at the cost of storing and processing padded entries. Grouped matrix multiplication (grouped GEMM) [5] offers another route by jointly scheduling matrix multiplications of different sizes without padding every sample to the largest edge count.

The relative benefits of these choices may depend on the dataset, particularly the distribution of graph sizes and edge counts, as well as batch size and hardware. We have added a discussion of this overhead and possible implementation optimizations to Appendix D, following Table 6. A systematic evaluation of these optimizations remains future work.

We sincerely appreciate your time and helpful suggestions. We welcome any further questions and would be happy to discuss these points in more detail.

[1] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning Mesh-Based Simulation with Graph Networks. In International Conference on Learning Representations, 2021.

[2] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B. Tenenbaum, and Antonio Torralba. Learning Particle Dynamics for Manipulating Rigid Bodies, Deformable Objects, and Fluids. In International Conference on Learning Representations, 2019.

[3] Jinwoo Kim, Tien Dat Nguyen, Seonwoo Min, Sungjun Cho, Moontae Lee, Honglak Lee, and Seunghoon Hong. Pure Transformers are Powerful Graph Learners. In Advances in Neural Information Processing Systems, 2022. Official batching implementation.

[4] Shilong Tao, Zhe Feng, Haonan Sun, Zhanxing Zhu, and Yunhuai Liu. Unisoma: A Unified Transformer-based Solver for Multi-Solid Systems. arXiv preprint arXiv:2506.06021, 2025.

[5] Babak Hejazi. Introducing Grouped GEMM APIs in cuBLAS and More Performance Updates. NVIDIA Technical Blog, 2024.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.