acceptodds
Under review as a conference paper at ICLR 2027

What Moves Together: Compositional Motion Modes for Physical Dynamics Prediction

Abstract

Physical dynamics prediction aims to forecast the trajectories of interacting components across diverse physical systems, from atoms in molecules to joints in human motion. The dynamics of such systems combine distinct but interacting kinds of motion: system-wide translation and rotation, system-wide deformation, and neighborhood-level motion. However, most existing approaches predict the resulting trajectory without explicitly separating these kinds of motion, leaving their individual contributions and interactions entangled. We introduce Compositional Mode Dynamics (CMD), a framework that represents physical motion with a compositional mode dictionary and models how its coefficients evolve and interact. The dictionary collects the spatial patterns of modes, grouped into three complementary families: rigid, collective, and local. Each mode follows its own linear dynamics, while structured coupling allows interactions within and across mode families without changing the underlying modal dynamics. This formulation yields an explicit and interpretable decomposition of predicted trajectories into physically meaningful motion components. Experiments on molecular dynamics and human motion datasets show that CMD achieves competitive prediction performance, while ablations and analyses validate the roles of the mode families and their coupled coefficient dynamics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
3D Physical dynamics, Representation learning, Motion decomposition, Compositional motion representation

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

The paper proposes a new model for a linear dynamical system which, combined with a neural network predicting the initial state of the system from initial observations, is applied to the prediction of future states of the modelled time series. The model itself is a mixture of handcrafted choices, forming an inductive bias, and learned parameters, which I describe below to the best of my understanding.

Up to normalisation and centring, the model is given by Xt=MKtθ0X_t = M K^t \theta_0 and θ0=fenc(X0,V0)\theta_0 = f_\text{enc}(X_0, V_0), where KK is the evolution matrix, MM the output matrix, KK the state-transition matrix, θt\theta_t the state, XtX_t the output, and fencf_\text{enc} is a graph neural network mapping the initial state (X0,V0)(X_0,V_0) (position and velocity) to the state.

The key innovation compared to just learning MM and KK is to give them a semi-analytical structure. So K=QΛQ⊤K = Q \Lambda Q^\top where QQ is a learned orthogonal matrix and Λ\Lambda is a block-diagonal matrix where each block is either capturing a damped oscillatory mode (with learnable frequency and decay parameters) or is a fixed "Jordan block" capturing non-oscillatory dynamics. These are mixed by the learned QQ to let the modes interact (without changing the eigenvalues of the systems, and hence the type of dynamics that they capture). A subset of the oscillatory modes are undamped and mapped to physical rotations of the object modelled.

MM is structured in a similar manner. For example, in the experiments XtX_t is a collection of 3D points (atoms in a molecule or joints in a human body). In this case, for the rotational mode above, the corresponding components of MM are set analytically using Rodrigues formulas, corresponding to an exact rotation of the 3D object. Other non-oscillatory components are mapped to exact xx, yy, and zz translations of the overall object. Other modes (such as the residual collective motion) are subspaces of the PCA space of the data, of dimension one or two, depending on whether they are associated to an oscillatory or non-oscillatory mode. These modes are defined globally and locally for groups of nearby points in XtX_t, given a mix of oscillatory and non-oscillatory modes, some global and some local, partially analytically defined and partially learned.

The learnable parts (CC, QQ, fencf_\text{enc} and frequencies and damping parameters) are fitted to data end-to-end on a training set of sequences. The model is then applied to a test set of new sequences, evaluated on predicting future XtX_t from the initial (X0,V0)(X_0, V_0). Experiments on small datasets of molecules and human poses show competitive performance compared to prior work in the area, ablations show the importance of the different modes considered in the model, and further analysis explores their interpretability.

Critical strengths
  • The paper appears to be technically sound.

  • In a world where inductive biases are all but removed from machine learning, this paper tries to introduce a set of sophisticated inductive biases to model complex 3D object dynamics.

  • The biases are interpretable, which allows the authors to diagnose, isolate, and correct specific problems with the model (e.g., in predicting the rotation component of the rotation-dominated motion of certain molecules).

Critical weaknesses
  • On a general level, it is very unclear that working on inductive biases like the ones proposed in the paper is a fruitful endeavour when data-driven general-purpose models have been shown over and over again to ultimately win out (e.g., famously, Sutton's "bitter lesson").

  • As it stands, I don't think the paper convincingly shows a specific advantage compared to more generic models. The proposed model is, in fact, quite complex, which has also the disadvantage of making it less practical for others to use. I would expect a model like this, which proposes a manually-defined system of interacting dynamics, to show a clear advantage over more general purpose solutions at least in some cases, and not just in terms of interpretability. The obvious example is when training data is very limited, where inductive biases could achieve better generalisation of approaches that do not have such biases, but this does not seem to be shown very directly in the experiment (see the questions below for more details).

  • The experimental data is quite limited. For example, in the human body pose sequences, the model is tasked with predicting just five (future time steps. While these are subsampled from a longer sequence of 30 frames, nevertheless this is not pushing the model capabilities much.

  • Results are also a little mixed. Despite focusing on small datasets and exploring complex hand-crafted inductive biases, the model does not uniformly wins out compare to prior work. Specifically, it loses to EGNO on the run sequences (i.e., 50% of the mocap data). This is imputed to a weakness of the proposed model which cannot adapt well to non-linear dynamics where, for instance, the rotation rate of limbs can change during the gait. This seems to be a significant limitation of this approach to modelling dynamic, particularly when more expressive models are available (and are often reliable).

  • The "interaction" model consists of rotating the state space. While this preserves the different dynamical modes, which is what the authors want, it also significantly restricts the "interactions" between these modes which, as far as I understand, are essentially additive in the output.

Questions for authors
  • Can the author provide experiments comparing their CDM with less-structured baselines (e.g,. just PCA + linear dynamics), showing a clear advantage when little training data is available and, presumably, less of an advantage when more training data is included?

  • Can the authors provide some evidence that the model has good generalisation capabilities? For example, if the model is trained on sequences where humans run slowly, would the model generalise to humans that run faster?

  • Can the authors show that the model has an advantage in long-horizon prediction? The model cam predict arbitrarily in the future quite easily and cheaply (just compute KtK^t, so why not increasing tt? Likewise, why subsampling time at all (except for being directly comparable to prior work)? Why not testing hundreds of time steps?

  • What kind of restrictions does using rotations of the state space to model "interactions" pose in practice?

10 comments
AuthorsOpenReview ↗

Response to Reviewer hzLz — Overview

We thank the reviewer for the detailed feedback. The central question is whether CMD’s structured inductive biases provide predictive benefits beyond interpretability (W1, W2, Q1).

Predictive benefits. Our additional experiments show that they do. CMD outperforms a free-dictionary variant across all ten systems (Section 1.2, W2), and the individual mode families each contribute to prediction accuracy (Section 1.3, W2). The advantage over less-structured baselines grows as training data increases (Section 1.1, W1, W2, Q1), and this trend persists in the substantially smaller CMD-lite (Section 1.1).

Temporal modelling. Beyond predictive accuracy at supervised time steps, CMD supports direct prediction at intermediate times without retraining (Section 2.1, Q3) and extrapolation beyond the training horizon without autoregressive rollout (Section 2.2, W3, Q3).

Limitations and expressivity. We also report where the structure does not help. A 1.53× shift in running speed remains challenging (Section 3, Q2), and the systems where CMD is not best are those dominated by rigid motion, where accuracy is limited by how the initial coefficients are estimated (Section 4, W4). Finally, we quantify both the benefits and the expressivity restrictions of cross-family coefficient coupling (Section 5, W5, Q4).

AuthorsOpenReview ↗

Response to Reviewer hzLz (1/7)


1. Predictive Benefits of Structured Dynamics (W1, W2, Q1)

1.1. Accuracy Across Training Data Sizes (W1, W2, Q1).

We vary the number of training samples over 100 / 250 / 500 / 1000; 500 is the setting of the paper. All models use the same data splits and evaluation protocol, and the neural models share the same epoch budget. PCA + linear is the baseline suggested in Q1: Cartesian PCA coefficients evolved by a single linear operator, without a neural network, with rank selected on validation. EGNO is retrained on the same splits. CMD-lite reduces our model's parameter count by 6-9$\times$, bringing it within 2.8-4.2$\times$ of EGNO.

Table 1. Test MSE (×10−2\times10^{-2}) versus the number of training samples.

Naphthalene

Model1002505001000
PCA + linear0.270.260.260.26
EGNO0.400.370.350.34
CMD0.310.260.220.20
CMD-lite0.320.260.230.20

Uracil

Model1002505001000
PCA + linear0.480.480.480.48
EGNO0.610.570.540.50
CMD0.440.390.340.30
CMD-lite0.440.390.340.30

Salicylic acid

Model1002505001000
PCA + linear0.570.570.560.56
EGNO1.321.110.740.70
CMD0.630.560.480.43
CMD-lite0.640.560.490.43

Toluene

Model1002505001000
PCA + linear4.624.514.474.46
EGNO4.884.884.864.84
CMD4.504.062.422.48
CMD-lite4.394.092.702.00

Ethanol

Model1002505001000
PCA + linear3.543.503.433.38
EGNO3.323.303.273.26
CMD3.393.122.692.41
CMD-lite3.403.092.712.56

Malonaldehyde

Model1002505001000
PCA + linear11.9811.8711.7311.71
EGNO11.0110.9010.8310.77
CMD12.0611.3910.7010.25
CMD-lite11.9111.3710.8110.15

Aspirin

Model1002505001000
PCA + linear7.547.337.267.24
EGNO7.737.377.036.91
CMD6.695.684.363.24
CMD-lite6.775.874.353.14

Benzene

Model1002505001000
PCA + linear3.142.972.882.84
EGNO30.9130.9130.8030.80
CMD3.162.511.701.33
CMD-lite2.692.542.221.41

The predictive advantage of CMD increases with training data. From 100 to 1000 samples, the geometric-mean MSE improves by 1.63$\times$ for CMD, compared with 1.16$\times$ for EGNO and 1.04$\times$ for PCA + linear. At 1000 samples, CMD attains 1.57$\times$ lower MSE than PCA + linear and 2.24$\times$ lower than EGNO. Both baselines show smaller improvements as data is added.

The same trend persists in CMD-lite. Despite a 6--9$\times$ reduction in parameters, it closely follows the full model's learning curve: from 100 to 1000 samples its geometric-mean MSE improves by 1.63$\times$, the same factor as the full model. The data-scaling behaviour is therefore not exclusive to the original high-capacity configuration.

At 100 samples CMD is within 1% of PCA + linear, so we do not claim a particular advantage in the lowest-data regime. Sections 1.2 and 1.3 examine the contribution of the structured dictionary and of the individual mode families.

AuthorsOpenReview ↗

Response to Reviewer hzLz (2/7)

1.2. Structured versus Free Dictionary (W2).

To evaluate the contribution of the structured dictionary, we train a free-dictionary variant whose columns are learned without predefined mode families or per-node organization. We retain the same encoder, coefficient-evolution formulation, auxiliary rigid branch, and training budget, with a matched dictionary-parameter budget. We first compare both models at CMD's learning rate and additionally tune the free dictionary's learning rate on Mocap.

Table 2. Structured versus free dictionary. Test MSE (×10−2\times10^{-2}; MD17 horizon 8, Mocap horizon 5).

SystemCMDFree (CMD LR)Free (tuned LR)
Naphthalene0.220.25---
Uracil0.340.35---
Salicylic acid0.480.53---
Toluene2.422.47---
Ethanol2.692.83---
Malonaldehyde10.7011.37---
Aspirin4.364.37---
Benzene1.704.06---
Mocap-Walk2.8950.205.85
Mocap-Run17.03136.6222.13

The free dictionary underperforms CMD on all ten systems. Across MD17, it has 1.17$\times$ higher geometric-mean MSE. On Mocap, the free dictionary exhibits training instability at CMD's learning rate. Even after tuning its learning rate, CMD maintains lower errors by 2.02$\times$ on Walk and 1.30$\times$ on Run. Using these tuned Mocap results, the free dictionary has 1.25$\times$ higher geometric-mean MSE across all ten systems.

These results demonstrate a consistent predictive advantage of the proposed structured representation beyond interpretability.

1.3. Per-Family Ablation (W2).

To assess the predictive contribution of each mode family, we remove one family at a time and retrain the model from scratch under the same training budget.

Table 3. Per-family ablation. Test MSE (×10−2\times10^{-2}; MD17 horizon 8, Mocap horizon 5).

SystemCMD−-Rigid−-Collective−-Local
Naphthalene0.220.230.520.29
Uracil0.340.370.780.40
Salicylic acid0.480.511.120.57
Toluene2.424.912.982.63
Ethanol2.693.332.692.66
Malonaldehyde10.7011.0010.7211.02
Aspirin4.365.287.334.32
Benzene1.7024.851.761.56
Mocap-Walk2.89459.54115.103.58
Mocap-Run17.033541.04465.0921.79

Removing the Rigid family increases prediction error in all ten systems, while removing Collective degrades performance in nine of ten. The effects are particularly pronounced on Mocap: removing Rigid increases MSE from 2.89 to 459.54 on Walk and from 17.03 to 3541.04 on Run, while removing Collective yields errors of 115.10 and 465.09, respectively.

Removing Local degrades performance in seven systems, with smaller and system-dependent effects. These results demonstrate the distinct predictive contributions of the three mode families beyond interpretability.

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

This paper proposes Compositional Mode Dynamics, which represents physical trajectories as combinations of rigid, collective, and local spatial modes. The model estimates mode coefficients from the initial state and evolves them through a linear operator with a prescribed spectral structure, allowing coupling between modes. An auxiliary rigid branch improves rigid coefficient estimation through coarse trajectory prediction. The authors evaluate forecasting performance on MD17 molecular dynamics and CMU human motion, using ablations and mode analyses to examine the components’ roles.

Critical strengths
  1. The modes are defined directly in physical displacement space, allowing predictions to be inspected by motion type. This structure provides a more intuitive analysis interface than general latent-state models.

  2. The method achieves competitive performance on several molecules and the Walk task. Auxiliary-branch ablations and initial-velocity timescale analyses identify specific forecasting bottlenecks, helping clarify the method’s applicability.

Critical weaknesses
  1. Rodrigues patterns can represent valid rotations, but unconstrained coefficients and cross-family coupling do not necessarily preserve rigidity.

  2. Overlapping motion families may explain the same displacement, leaving the physical interpretation of the decomposition ambiguous.

  3. Existing ablations examine coupling and the auxiliary branch but do not isolate the benefit of organizing modes into three families.

Questions for authors
  1. Do rigid-only predictions preserve pairwise distances? If not, how large are the deviations?

  2. How is motion attributed when families overlap? Could controlled experiments with known motion components validate the decomposition?

  3. How does an equally sized learned mode dictionary without predefined families perform? What happens when each family is removed?

8 comments
AuthorsOpenReview ↗

Response to Reviewer RRLF — Overview

We thank the reviewer for the detailed feedback. The concerns fall into two groups, and we address both with new analysis and new experiments.

On structural necessity (W3, Q3), we add two experiments. We train a free-dictionary variant that removes the three-family organization while holding the encoder, the coefficient dynamics, the auxiliary rigid branch, a matched dictionary-parameter budget and the training budget fixed, and we retrain the model from scratch with each family removed in turn. The first comparison measures what the structured dictionary contributes, while the ablations quantify the predictive contribution of each family.

On physical validity and attribution (W1, W2, Q1, Q2), we add measurements that go beyond the contribution magnitudes of the original submission. We evaluate whether each family's output - after learned coefficient evolution and cross-family coupling - retains its intended geometric or spatial properties, including a direct measurement of pairwise-distance deviations in rigid-only predictions. We then quantify how an observed motion component is distributed across the three families and how much the families overlap or cancel.

AuthorsOpenReview ↗

Response to Reviewer RRLF (1/4)


1. Physical Significance and Validity (W1, W2, Q1, Q2)

1.1. Family-Specific Physical Properties (W1, Q1). For each family ff, let uf∈RN×3\mathbf{u}^{f}\in\mathbb{R}^{N\times3} denote the per-atom displacement over the prediction horizon contributed by that family's dictionary columns alone.

We evaluate whether the learned family outputs exhibit the geometric or spatial properties intended by their dictionary design. The "By design" column provides a reference measured from the constructed dictionary columns before learned coefficient evolution. The "No coupling" and "Coupled" columns report the corresponding properties of the predicted family outputs without and with cross-family coefficient interactions, respectively.

Rigid. We measure the root-mean-square change in pairwise distances under urigid\mathbf{u}^{\mathrm{rigid}}, denoted by εrigid\varepsilon_{\mathrm{rigid}}. The "By design" column evaluates the first-order rotation columns at the observed rotation angle per horizon; translations preserve pairwise distances exactly. The "Observed" column gives the ground-truth deformation as a reference scale, not as a bound on rigidity. All distance-deviation values are reported in Å for MD17 and dataset coordinate units for Mocap; θ\theta is measured in degrees.

Table 1. Rigid-family pairwise-distance deviation, ordered by the observed rotation per horizon.

Systemθ\theta/horizonBy designNo couplingCoupledObserved
Naphthalene0.8∘0.8^\circ0.00020.0010.0010.074
Uracil1.2∘1.2^\circ0.00040.0010.0020.077
Salicylic acid1.2∘1.2^\circ0.00050.0010.0030.095
Toluene6.7∘6.7^\circ0.0130.0070.0110.072
Aspirin9.0∘9.0^\circ0.0310.0080.0750.242
Malonaldehyde9.1∘9.1^\circ0.0230.0800.0610.254
Mocap-Walk9.1∘9.1^\circ0.1150.0550.3621.225
Mocap-Run16.7∘16.7^\circ0.3551.7110.8262.655
Ethanol18.7∘18.7^\circ0.0700.0800.0630.073
Benzene64.7∘64.7^\circ0.8235.9961.9030.040

For ideal rigid-body motion, pairwise distances remain unchanged. The observed deformation provides a reference scale for interpreting the deviations: with coupling, the rigid family exhibits less internal deformation than the actual motion in 9 of 10 systems, although this does not imply exact rigidity.

Cross-family coupling can either increase or decrease the rigid-family distance deviation, rather than uniformly enforcing rigidity.

The exception is Benzene, which has the largest observed rotation (64.7∘64.7^\circ per horizon). Coupling reduces Benzene's deviation from 5.996 to 1.903 but does not eliminate it.

Thus, in response to Q1, the rigid family provides a geometric inductive bias rather than an exact rigidity constraint. Since its coefficients are unconstrained and cross-family coupling is allowed, exact rigidity is not guaranteed. As further examined in Section 1.2, Benzene also exhibits opposing rigid and local contributions along the reference rigid direction (+1.416 and −0.420-0.420, respectively).

AuthorsOpenReview ↗

Response to Reviewer RRLF (2/4)

Collective. We measure the spatial extent of collective motion using the effective number of participating atoms:

\frac{ \big(\sum_i \lVert\mathbf{u}_i^{\mathrm{col}}\rVert^2\big)^2 }{ \sum_i \lVert\mathbf{u}_i^{\mathrm{col}}\rVert^4 }.$$ The value ranges from 1 (one moving atom) to $N$ (equal displacement magnitudes across all atoms). The *"By design"* value is obtained by computing this metric for each collective dictionary column individually and averaging across columns. The *"No coupling"* and *"Coupled"* values are computed from the realized collective-family displacements after coefficient evolution. **Table 2.** Effective number of participating atoms in the collective family. | System | Ideal ($N$) | By design | No coupling | Coupled | | --- | ---: | ---: | ---: | ---: | | Naphthalene | 10 | 5.9 | 6.2 | 5.5 | | Uracil | 8 | 4.3 | 3.7 | 3.9 | | Salicylic acid | 10 | 5.3 | 5.3 | 5.0 | | Toluene | 7 | 4.5 | 4.3 | 4.3 | | Ethanol | 3 | 2.3 | 2.3 | 2.4 | | Malonaldehyde | 5 | 3.4 | 3.6 | 3.6 | | Aspirin | 13 | 5.2 | 4.6 | 4.4 | | Benzene | 6 | 4.9 | 4.9 | 5.0 | | Mocap-Walk | 31 | 11.2 | 7.6 | 7.0 | | Mocap-Run | 31 | 12.4 | 9.7 | 10.0 | The collective family produces spatially distributed displacements across all systems. The realized effective number of participating atoms ranges from 2.4 to 5.5 for MD17 and reaches 10.0 for Mocap-Run and 7.0 for Mocap-Walk, indicating that the predicted motion involves multiple atoms rather than being concentrated on a single atom. This spatially distributed behavior is also consistent with the dictionary design. In the seven smallest systems, the *"Coupled"* values remain within 10% of their respective *"By design"* values. For the three largest systems (Aspirin, Mocap-Run, and Mocap-Walk), they retain 85%, 81%, and 63% of the design-level effective participation, respectively. Thus, while the extent of participation varies across systems, the collective family exhibits the intended spatially distributed behavior after coefficient evolution. **Local.** We evaluate whether the Local family produces spatially localized displacements using two-hop localization enrichment. For each node, this metric is defined as the fraction of its local-column displacement energy within its two-hop neighborhood, divided by the fraction of nodes belonging to that neighborhood. A value greater than 1 indicates that displacement energy is preferentially concentrated within the two-hop neighborhood. The *"Ideal"* value corresponds to complete concentration within the neighborhood. The *"By design"* value is computed from the constructed local dictionary columns before learned coefficient evolution. The *"No coupling"* and *"Coupled"* values are computed from the corresponding realized local-family outputs. The metric is evaluated separately for each node's local columns and then averaged across nodes. **Table 3.** Two-hop localization enrichment of the local family, ordered by decreasing ideal value. | System | $N$ | 2-hop size | Ideal | By design | No coupling | Coupled | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | Mocap-Walk | 31 | 5.2 | 5.97 | 4.56 | 4.31 | 4.47 | | Mocap-Run | 31 | 5.2 | 5.97 | 4.50 | 3.82 | 4.31 | | Aspirin | 13 | 5.6 | 2.32 | 1.42 | 1.42 | 1.41 | | Salicylic acid | 10 | 5.6 | 1.79 | 1.16 | 1.02 | 1.15 | | Naphthalene | 10 | 6.0 | 1.67 | 1.13 | 1.07 | 1.12 | | Uracil | 8 | 5.5 | 1.45 | 1.01 | 0.95 | 0.97 | | Toluene | 7 | 5.3 | 1.32 | 0.96 | 0.97 | 0.97 | | Malonaldehyde | 5 | 3.8 | 1.32 | 1.02 | 1.12 | 1.04 | | Benzene | 6 | 5.0 | 1.20 | 0.91 | 0.90 | 0.92 | | Ethanol | 3 | 3.0 | 1.00 | 1.00 | 1.00 | 1.00 | The intended locality is most clearly observed in Mocap. The realized enrichment reaches 4.47 for Walk and 4.31 for Run, compared with an ideal value of 5.97. These results indicate that displacement energy is preferentially concentrated within each node's two-hop neighborhood, consistent with the intended behavior of the Local family. For molecular systems, two-hop neighborhoods cover a larger fraction of the atoms, limiting the metric's ability to distinguish localized from distributed motion. Consequently, the results provide stronger evidence of the intended locality in Mocap than in small molecules. Together, these analyses evaluate whether the learned family outputs exhibit their intended geometric and spatial properties, beyond their contribution magnitudes. The rigid family generally introduces limited internal deformation, the collective family produces spatially distributed displacements, and the local family exhibits clear neighborhood localization, particularly in Mocap. These findings support the intended behaviors of the three families while also identifying the limited discriminative power of the two-hop localization metric in small molecules.
Reviewer_RNQGOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses physical dynamics prediction, which aims to forecast the trajectories of interacting components from an initial state. It introduces Compositional Mode Dynamics (CMD), a framework that represents physical motion with a compositional mode dictionary organized into three mode families: rigid, collective, and local. CMD estimates initial mode coefficients from the initial state and evolves them through structured coupled coefficient dynamics, CMD is evaluated on MD17 and CMU Motion Capture through comparisons with baselines, ablations of mode coupling and the auxiliary rigid branch, and analyses of the coefficient space and per-family decomposition of predicted motion.

Critical strengths

The model offers a clear way to examine its predictions. It employs three families of modes—rigid, collective, and local—coupling their coefficients while preserving the modal spectrum, with the contribution of each family remaining independently observable. I find this alignment between model design and output highly valuable; it is, in fact, the primary reason I appreciate this paper.

The model performs exceptionally well across several tasks. It achieves the lowest prediction error for four out of the eight molecules in the MD17 dataset, as well as the lowest error on the Walk task from the CMU Motion Capture dataset. These results demonstrate that the framework is fully capable of competing with strong baselines.

Critical weaknesses

However, the physical significance of this decomposition requires further substantiation. While the plots and contribution scores illustrate what each mode family predicts, do these components truly capture the specific types of motion they are intended to represent? For instance, given the learned coefficients and couplings, does the rigid component actually maintain its shape? As this decomposition constitutes one of the paper's core contributions, I require more compelling evidence to support its acceptance.

The ablation studies fail to clearly quantify the specific gains provided by the "three-mode-family" structure. Although the experiments examine the coupling mechanism and the auxiliary rigid branch, they do not isolate the impact of each individual mode family. Furthermore, I would like to see a comparison against a learned dictionary model that lacks this three-mode-family structure but utilizes an equivalent encoder and a comparable number of parameters. Without such tests, it is difficult to determine to what extent the reported performance improvements can be attributed to the model's core design. This is the primary reason I am inclined to recommend rejection. In addition, the experiments in Table 2 did not show much competitiveness, and the number of decimal places in Table 3 was not standardized.

Questions for authors
  1. Do these three mode families accurately capture the types of motion they are intended to represent? The paper’s central contribution lies in decomposing motion into rigid, collective, and local components. My question is: is there sufficient evidence to confirm that these components truly capture the specific motion types they are meant to correspond to? (1) Beyond the visualization and quantitative metrics for the contribution of each mode, what other quantitative evidence exists to validate the effectiveness of this decomposition? Contribution metrics merely reflect the magnitude (amplitude) of each component and do not directly prove that the motions have been correctly classified. (2) Given the potential for overlap between different mode families, how do you assess whether a motion belonging to one category has been misclassified into another?

  2. To what extent can the improvement in predictive performance be attributed to this "three-mode-family" structure? The current ablation studies primarily examine the coupling configuration and the auxiliary rigid-body branch. My question is: does this three-mode-family structure itself offer unique advantages compared to other representation schemes? (1) How does predictive accuracy change when rigid-body modes, collective modes, or local modes are removed individually? Can these results clarify the specific roles of each mode family across different tasks, and help distinguish the impact of the "mode structure itself" from the effects of a reduction in "model capacity"? (2) Keeping the encoder, coefficient dynamics, and auxiliary rigid-body branch constant, how does CMD perform compared to a mode dictionary learned without the proposed family rules? Such a comparison would help isolate the contribution of the physical organization of the modes from the contributions of other components and the overall model capacity.

7 comments
AuthorsOpenReview ↗

Response to Reviewer RNQG — Overview

We thank the reviewer for the detailed feedback. We address the concerns about physical significance and structural necessity with additional analyses and experiments.

On physical significance (W1, Q1), we measure whether the outputs of the three families retain their intended geometric or spatial properties after coefficient evolution and cross-family coupling. We also quantify how observed motion components are attributed to the predicted families and measure their overlap and cancellation.

On structural necessity (W2, Q2), we compare CMD with a free learned dictionary lacking family and per-node organization, and retrain the model from scratch with each family removed in turn. The first comparison examines the effect of dictionary organization, while the second quantifies the predictive contribution of each family.

We additionally address benchmark competitiveness (W2) and numerical formatting (W2).

AuthorsOpenReview ↗

Response to Reviewer RNQG (1/4)


1. Physical Significance and Validity (W1, Q1)

1.1. Family-Specific Physical Properties (W1, Q1)

For each family ff, let uf∈RN×3\mathbf{u}^{f}\in\mathbb{R}^{N\times3} denote the per-atom displacement over the prediction horizon contributed by that family's dictionary columns alone.

We evaluate whether the learned family outputs exhibit the geometric or spatial properties intended by their dictionary design. The "By design" column provides a reference measured from the constructed dictionary columns before learned coefficient evolution. The "No coupling" and "Coupled" columns report the corresponding properties of the predicted family outputs without and with cross-family coefficient interactions, respectively.

Rigid. We measure the root-mean-square change in pairwise distances under urigid\mathbf{u}^{\mathrm{rigid}}, denoted by εrigid\varepsilon_{\mathrm{rigid}}. The "By design" column evaluates the first-order rotation columns at the observed rotation angle per horizon; translations preserve pairwise distances exactly. The "Observed" column gives the ground-truth deformation as a reference scale, not as a bound on rigidity. All distance-deviation values are reported in Å for MD17 and dataset coordinate units for Mocap; θ\theta is measured in degrees.

Table 1. Rigid-family pairwise-distance deviation, ordered by the observed rotation per horizon.

Systemθ\theta/horizonBy designNo couplingCoupledObserved
Naphthalene0.8∘0.8^\circ0.00020.0010.0010.074
Uracil1.2∘1.2^\circ0.00040.0010.0020.077
Salicylic acid1.2∘1.2^\circ0.00050.0010.0030.095
Toluene6.7∘6.7^\circ0.0130.0070.0110.072
Aspirin9.0∘9.0^\circ0.0310.0080.0750.242
Malonaldehyde9.1∘9.1^\circ0.0230.0800.0610.254
Mocap-Walk9.1∘9.1^\circ0.1150.0550.3621.225
Mocap-Run16.7∘16.7^\circ0.3551.7110.8262.655
Ethanol18.7∘18.7^\circ0.0700.0800.0630.073
Benzene64.7∘64.7^\circ0.8235.9961.9030.040

For ideal rigid-body motion, pairwise distances remain unchanged. The observed deformation provides a reference scale for interpreting the deviations: with coupling, the rigid family exhibits less internal deformation than the actual motion in 9 of 10 systems, although this does not imply exact rigidity.

The exception is Benzene, which has the largest observed rotation (64.7∘64.7^\circ per horizon). Coupling reduces Benzene's deviation from 5.996 to 1.903 but does not eliminate it. Thus, rigidity is an inductive bias rather than an exact constraint.

AuthorsOpenReview ↗

Response to Reviewer RNQG (2/4)

Collective. We measure the spatial extent of collective motion using the effective number of participating atoms:

\frac{ \big(\sum_i \\|\mathbf{u}_i^{\mathrm{col}}\\|^2\big)^2 }{ \sum_i \\|\mathbf{u}_i^{\mathrm{col}}\\|^4 }.$$ The value ranges from 1 (one moving atom) to $N$ (equal displacement magnitudes across all atoms). The *"By design"* value is obtained by computing this metric for each collective dictionary column individually and averaging across columns. The *"No coupling"* and *"Coupled"* values are computed from the realized collective-family displacements after coefficient evolution. **Table 2.** Effective number of participating atoms in the collective family. | System | Ideal ($N$) | By design | No coupling | Coupled | | --- | ---: | ---: | ---: | ---: | | Naphthalene | 10 | 5.9 | 6.2 | 5.5 | | Uracil | 8 | 4.3 | 3.7 | 3.9 | | Salicylic acid | 10 | 5.3 | 5.3 | 5.0 | | Toluene | 7 | 4.5 | 4.3 | 4.3 | | Ethanol | 3 | 2.3 | 2.3 | 2.4 | | Malonaldehyde | 5 | 3.4 | 3.6 | 3.6 | | Aspirin | 13 | 5.2 | 4.6 | 4.4 | | Benzene | 6 | 4.9 | 4.9 | 5.0 | | Mocap-Walk | 31 | 11.2 | 7.6 | 7.0 | | Mocap-Run | 31 | 12.4 | 9.7 | 10.0 | The collective family produces spatially distributed displacements across all systems. The realized effective number of participating atoms ranges from 2.4 to 5.5 for MD17 and reaches 10.0 for Mocap-Run and 7.0 for Mocap-Walk, indicating that the predicted motion involves multiple atoms rather than being concentrated on a single atom. This spatially distributed behavior is also consistent with the dictionary design. In the seven smallest systems, the *"Coupled"* values remain within 10% of their respective *"By design"* values. For the three largest systems (Aspirin, Mocap-Run, and Mocap-Walk), they retain 85%, 81%, and 63% of the design-level effective participation, respectively. Thus, while the extent of participation varies across systems, the collective family exhibits the intended spatially distributed behavior after coefficient evolution. **Local.** We evaluate whether the Local family produces spatially localized displacements using two-hop localization enrichment. For each node, this metric is defined as the fraction of its local-column displacement energy within its two-hop neighborhood, divided by the fraction of nodes belonging to that neighborhood. A value greater than 1 indicates that displacement energy is preferentially concentrated within the two-hop neighborhood. The *"Ideal"* value corresponds to complete concentration within the neighborhood. The *"By design"* value is computed from the constructed local dictionary columns before learned coefficient evolution. The *"No coupling"* and *"Coupled"* values are computed from the corresponding realized local-family outputs. The metric is evaluated separately for each node's local columns and then averaged across nodes. **Table 3.** Two-hop localization enrichment of the local family, ordered by the decreasing ideal value. | System | $N$ | 2-hop size | Ideal | By design | No coupling | Coupled | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | Mocap-Walk | 31 | 5.2 | 5.97 | 4.56 | 4.31 | 4.47 | | Mocap-Run | 31 | 5.2 | 5.97 | 4.50 | 3.82 | 4.31 | | Aspirin | 13 | 5.6 | 2.32 | 1.42 | 1.42 | 1.41 | | Salicylic acid | 10 | 5.6 | 1.79 | 1.16 | 1.02 | 1.15 | | Naphthalene | 10 | 6.0 | 1.67 | 1.13 | 1.07 | 1.12 | | Uracil | 8 | 5.5 | 1.45 | 1.01 | 0.95 | 0.97 | | Toluene | 7 | 5.3 | 1.32 | 0.96 | 0.97 | 0.97 | | Malonaldehyde | 5 | 3.8 | 1.32 | 1.02 | 1.12 | 1.04 | | Benzene | 6 | 5.0 | 1.20 | 0.91 | 0.90 | 0.92 | | Ethanol | 3 | 3.0 | 1.00 | 1.00 | 1.00 | 1.00 | The intended locality is most clearly observed in Mocap. The realized enrichment reaches 4.47 for Walk and 4.31 for Run, compared with an ideal value of 5.97. These results indicate that displacement energy is preferentially concentrated within each node's two-hop neighborhood, consistent with the intended behavior of the Local family. For molecular systems, two-hop neighborhoods cover a larger fraction of the atoms, limiting the metric's ability to distinguish localized from distributed motion. Consequently, the results provide stronger evidence of the intended locality in Mocap than in small molecules.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all Reviewers

We sincerely thank all reviewers for their thoughtful feedback and constructive suggestions. Their comments helped us strengthen the physical and predictive justification of CMD and motivated an important revision to the model itself.

We are also encouraged that the reviewers highlighted the physically interpretable mode decomposition and independently observable family contributions (Reviewers RNQG, RRLF, hzLz), competitive prediction performance on MD17 and Mocap-Walk (Reviewers RNQG, RRLF), and the ability to diagnose specific forecasting bottlenecks through interpretable inductive biases (Reviewers RRLF, hzLz).

 

Model Revision: Extended Orthogonalization of the Local Dictionary.

In response to concerns about overlapping mode families (Reviewers RNQG, RRLF), we extended the existing orthogonalization of the local dictionary. The local columns were already projected to be orthogonal to the global rigid subspace of translations and first-order rotations; we now extend this projection to the collective subspace, so that local and collective dictionary columns are orthogonal by construction. This reduces ambiguity in the motion decomposition while preserving predictive accuracy. All results in the revised manuscript and individual responses use this refined construction. The appendix compares both constructions.

 

Additional Experiments and Analyses.

Beyond this revision, we conducted new experiments and analyses addressing three main concerns.

  1. Physical validity and interpretability (Reviewers RNQG, RRLF). We provide quantitative evidence that the three mode families exhibit their intended physical properties and capture the corresponding types of motion, while identifying remaining ambiguities.

  2. Structural necessity and predictive benefits (Reviewers RNQG, RRLF, hzLz). We demonstrate that the structured representation contributes to predictive accuracy beyond interpretability, supported by comparisons with less-structured alternatives and targeted ablations.

  3. Predictive capabilities and limitations (Reviewer hzLz). We extend the evaluation beyond the standard prediction protocol and investigate generalisation, prediction bottlenecks, and expressivity limitations to clarify where CMD succeeds and where it remains challenging.

Detailed results and discussions are provided in our individual responses. We appreciate the opportunity to further clarify our contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.