acceptodds
Withdrawn from ICLR 2027

Geometric Invariance of Gradient-Induced Coupling for Continuous 3D Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We develop an -equivariant graph neural network with continuous latent dynamics, using a shared continuous vector field across nodes and trajectories. Motivated by the regularities shared across diverse physical states, we use the same field across latent states. Because the same field is evaluated at different latent states, gradient updates induce cross-state coupling through the shared-field parameters. We characterize this gradient-induced cross-state coupling and introduce a geometric symmetry bias that makes it invariant under common latent transformations and dependent on relative latent configuration. Our experiments show that invariant coupling can be reflected in latent organization. Across 3D dynamics benchmarks, the continuous formulation remains competitive with established baselines in forecasting performance, even under random irregular temporal sampling. Importantly, the framework unifies continuous latent dynamics with explicit geometric control over cross-state coupling without sacrificing physical equivariance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
geometric deep learning, E(3)-equivariance, continuous dynamics, latent dynamics, neural tangent kernels, Lie groups

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_Y4G7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors develop a 3D motion prediction model by combining an existing equivariant graph neural network with a continuous latent dynamical system. Their main contribution is the design and comparison of three parameterisations of the shared latent vector field: polynomial, Fourier, and SU(2) harmonic bases.

The authors evaluate whether these parameterisations produce better-organised latent representations and more accurate physical predictions. Experiments show some improvements in latent organisation, but these vary across datasets and do not consistently improve forecasting accuracy.

Critical strengths
  • The authors provide a clear mathematical characterisation of the parameter-gradient kernels induced by the three latent vector-field parameterisations, although these results largely follow from established kernel and representation theory.
  • The authors compare models with similar numbers of latent-field parameters across controlled physical systems and established dynamics benchmarks.
Critical weaknesses

The central motivation lacks a justified mechanism. There is no established reason why symmetry of learning influence should improve the correspondence between latent distances and physical variables or temporal neighbourhoods. The experiments do not isolate this proposed mechanism. Moreover, the results are inconsistent with a general benefit from kernel invariance. For example, the non-invariant polynomial model achieves better relational temporal locality than both invariant models on N-body.

The theoretical analysis does not characterise the implemented learning process or establish the role of the latent dynamics. The kernel analysis describes coefficient updates under ordinary gradient descent with fixed latent inputs, while the actual models use Adam and jointly train the encoder and decoder. The authors do not establish whether the claimed invariance holds for the realised parameter updates. Furthermore, the encoder–flow–decoder architecture does not enforce correspondence between physical and latent flows, and the decoder receives a separate physical prediction pathway. No flow-consistency evaluation or no-flow baseline establishes whether the latent ODE learns meaningful physical dynamics or merely generates useful features for forecasting. These gaps substantially weaken the interpretation of the theoretical and empirical contributions.

Questions for authors
  1. The derivation assumes gradient descent on the shared-field coefficients, but the models are trained end-to-end using Adam. Since adaptive parameter updates can change the effective coupling kernel, what evidence shows that the proposed invariance describes the actual learning process?
  2. The theory establishes invariance of the parameter-gradient kernel, while the experiments measure correlations with physical variables and temporal neighbourhoods. What mechanism connects these properties?
  3. The decoder receives physical information through a separate spatial pathway, and the architecture does not enforce correspondence between physical and latent flows. What evidence establishes that the latent ODE learns meaningful dynamical evolution, rather than acting as a time-dependent feature generator?
Reviewer_k7UtOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces continuous vector fields to model the continuous latent dynamics in 3D interacting systems, and integrate them in equivariant graph networks. The authors explore three different parameterizations of the field, ranging from non-invariant to fully invariant.

Experiments in common benchmarks in dynamical systems demonstrate the effectiveness of the method in terms of forecasting, while controlled dynamics experiments show that geometri inductive biases can be beneficial for cross-state learning.

Critical strengths

The notion of gradient-induced cross-state coupling is interesting and potentially useful for modelling the dynamics of dynamical systems.

Critical weaknesses

The paper is, unfortunately, not very well written and thus, it is hard to follow. The source code is not shared, which makes it difficult to assess the validity and the effectiveness of the proposed method practically.

The experiments are performed on common datasets in the field, but these datasets are quite saturated and small-scale by today's standards. Hence, it is hard to evaluate whether the proposed method scales to larger graphs and dynamical systems. Furthermore, with the exception of EGNO, the baselines used are rather dated.

Questions for authors
  • Would an MLP-based vector field be a useful field parameterization? Some experiments show that coupling invariance is not always necessary.
  • I would expect a comparison with the continuous time dynamics models referenced in the introduction. Can the cross-state coupling be observed in these works?

Minor

Some references are incorrect, e.g. Knapp & Knapp, 1996, instead of Knapp, 1996.

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

This paper proposes a modular framework where an EGNN backbone encodes physical states into E(3)-invariant latent representations, which are then evolved by a shared continuous vector field, and decoded back to equivariant physical predictions. The core intellectual contribution is characterizing the “gradient-induced cross-state coupling” (repeated continuously through out the paper more than 10 times!) that arises when a shared field is evaluated at different latent states, and showing that the symmetry of the resulting parameter-gradient kernel depends on the choice of latent geometry and basis. Three realizations are compared: R3-Poly (polynomial, generally non-invariant kernel), R3-Fourier (Fourier, translation-invariant kernel), and SU2-PW (Peter-Weyl on SU(2), invariant under common group actions). The paper is carefully written, mathematically precise, and honest about what it achieves.

Critical strengths

The mathematical development seems to be presented correctly. The kernel derivations for all three realizations are presented elegantly. The paper is honest about its negative results, which was refreshing to see. The benchmark results section acknowledges that forecasting performance doesn’t track the coupling properties.

The three-way comparison is well designed for isolating factors. R3-Poly vs R3-Fourier isolates cross-state invariance within the same latent geometry; R3-Fourier vs SU2-PW isolates the effect of the geometry itself while both maintain invariance. This factorial structure is what makes the controlled results interpretable.

The continuous formulation handles irregular temporal sampling without degradation. Tables 3, 4, and 6 show that random irregular sampling produces comparable F-MSE and A-MSE to regular sampling across all benchmarks. This is a practical capability that discrete-step models don’t have, and it’s demonstrated in the results.

The controlled experiments directly test the hypothesis presented. The authors construct systems with known physical variables (acceleration magnitude, subsystem mass) and measure physical-latent organization with Spearman and distance correlation. The Impulse control in particular was well designed for exposing whether within-state relational geometry tracks physical structure.

Critical weaknesses

On every benchmark, the three realizations perform comparably to each other and to the EGNO baseline in forecasting, implying perhaps that the entire latent-dynamics and kernel-symmetry machinery provides no measurable downstream benefit? The hypothesis that invariant coupling produces better latent organization is confirmed only on toy controlled systems and is explicitly “system-dependent” on the real benchmarks, where D-PR and G-PR show no consistent ordering and R3-Poly achieves H = 0 on Benzene despite being theoretically non-invariant. The latent dimension is locked at 3 for all realizations with no ablation, and the integrators differ between the R3 and SU(2) realizations (Euler vs exponential map), introducing constraints that are never controlled for.

Questions for authors
  1. Can the authors clarify the kernel analysis from standard kernel-method theory? Since the shared fields are linear in parameters, the cross-state kernel is parameter-independent and equals the Gram matrix of the feature map — the same object that governs ridge regression, Bayesian linear regression, or any linear update rule on these features. What does the “gradient-induced” framing add beyond a re-derivation of this fact?

  2. It would be good to see if we can see some ablation studies of the R3-Poly normalization against an unnormalized polynomial, or use the same normalization for all three realizations? The current comparison confounds basis symmetry with a capacity constraint (constant field magnitude) specific to R3-Poly, and without this control the Impulse result could reflect the normalization rather than the non-invariance.

  3. Does varying the latent dimension for the R3 realizations change the organization gap with SU2-PW? What is the null model for D-PR and G-PR?

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

Related papers

Discussion (0)

Sign in to comment.