acceptodds
Under review as a conference paper at ICLR 2027

Geometry Is Not Enough: Symplectic Structure and Noise Robustness in JEPA World Models for Beam Dynamics

Abstract

Joint-Embedding Predictive Architectures (JEPAs) learn dynamics in latent space and promise cheaper, noise-robust learning-based surrogates for physical simulation. Without a reconstruction loss anchoring the representation, however, the encoder and predictor are free to settle on latents that are mutually consistent with little knowledge about the system dynamics. But whether enforcing that structure in latent space improves performance remains poorly understood. We address this by studying JEPAs on a system whose latent dynamics have a known target: longitudinal beam dynamics, whose one-turn map is a symplectic Hamiltonian flow, and use that structure as a reference for the learned latent. Our contributions are threefold: we examine whether enforcing symplectic structure on the latent predictor restores the geometry of the true one-turn map and improves accuracy, benchmark JEPA against reconstruction-based and pixel-space surrogates in error and cost, and assess its robustness to noise and distribution shift. On BLonD simulations of the FNAL Recycler, we find that symplectic predictors restore the geometry of the one-turn map but only direct map parameterisations improve accuracy, and that JEPA trains – faster, and filters input noise, though reconstruction retains the lowest error. Our findings align with prior work on transfer-operator and latent Hamiltonian learning and point to encoder objectives and training regimes that would make JEPA a reliable surrogate for complex physical dynamics.

Primary area:
learning on time series and dynamical systems
Keywords:
Joint-Embedding Predictive Architectures, symplectic structure preservation, learned surrogates, beam dynamics, particle accelerators

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

This paper investigates whether Joint-Embedding Predictive Architectures (JEPAs) can serve as efficient and reliable surrogate models for physical dynamical systems, particularly longitudinal particle beam dynamics. The authors focus on three questions: whether imposing symplectic structure improves prediction, how JEPA compares with reconstruction-based models in accuracy and computational cost, and how these models perform under observation noise and out-of-distribution data. The evaluation results show that imposing symplectic structure improves the geometric properties of latent transitions, but does not always improve prediction accuracy. JEPA achieves faster training and better robustness to additive Gaussian noise, while reconstruction-based models generally achieve better prediction accuracy under clean observations.

Critical strengths
  1. I appreciate that the authors do not simply show that enforcing symplecticity improves prediction. Instead, they compare multiple architectures and investigate why some structure-preserving models perform better than others. For example, Figure 2 shows that HNN and SRNN satisfy the geometric constraints but still perform worse than the unconstrained baseline. I think this is an interesting observation, as it suggests that imposing the correct physical structure on the predictor is not sufficient if the learned latent coordinates are not suitable.

  2. The comparison between JEPA and reconstruction-based models is interesting. Although the autoencoder achieves better accuracy on clean data, JEPA is much less affected by additive Gaussian noise. I also appreciate that the authors investigate Poisson noise and OOD data, where the conclusions are different. These experiments help explain both the advantages and limitations of the JEPA objective rather than presenting it as consistently superior.

Critical weaknesses
  1. My main concern is that the paper directly imposes symplectic constraints on the 64-dimensional latent space based on the symplectic property of the original two-dimensional particle dynamics. However, it is unclear whether this property should still hold after an arbitrary nonlinear encoding of the phase-space density. Symplecticity is preserved under canonical coordinate transformations, but the CNN encoder is not constrained to learn such transformations. Therefore, even if a latent predictor satisfies the symplectic conditions in Figure 2, this does not necessarily mean that it preserves the actual physical structure of the beam dynamics. I think this is important because one of the main contributions is to examine whether the physical geometry can be restored through structural constraints. The authors should better justify the connection between the original physical symplecticity and the imposed latent constraints.

  2. In Section 3, the authors assume that the beam dynamics follow a deterministic and autonomous one-turn map, where the current state uniquely determines the next state. However, Appendix B shows that the RF voltages are sampled independently over 60–100 kV across simulations. Since these machine settings affect the synchrotron dynamics, different simulations may have different transition maps even when their current beam distributions are similar. Meanwhile, the learned predictors only take the encoded density frame as input, without explicitly including the machine settings. I wonder whether this introduces ambiguity into the learning problem, since the model is effectively asked to approximate multiple dynamical systems using a single transition function. This may also affect the comparison between different predictor architectures, particularly the linear and Hamiltonian models. I think the authors should clarify this assumption and investigate whether conditioning on machine parameters changes the conclusions.

  3. One concern is the choice of VISReg as the default regularizer. In Appendix C, Table 4, the authors show that VICReg performs considerably better than VISReg. For example, VICReg reduces the mean rollout MSE ratio from 0.16 to 0.08 and the final-step MSE ratio from 0.69 to 0.33. Interestingly, this result is already very close to the Linear-orth Autoencoder in Table 1, which achieves 0.07 and 0.31, respectively. However, the main comparisons are still conducted using VISReg. The authors explain their preference for VISReg, but I am not fully convinced that this is sufficient, since the paper aims to evaluate the performance trade-offs between JEPA and reconstruction-based models. I think using a weaker JEPA configuration may affect the fairness of the comparison, and the authors should at least report the main accuracy and noise-robustness results using VICReg as well.

  4. The paper emphasizes long-horizon prediction as an important motivation, especially for physical simulations that need to track beam dynamics over thousands of turns. However, Section 5 states that each simulation contains only 250 turns, corresponding to about 0.88 of one nominal synchrotron period. Therefore, even the longest evaluation does not cover a complete oscillation cycle. I wonder whether the reported performance can be maintained over multiple synchrotron periods, where small errors in oscillation frequency and phase may accumulate significantly. This is particularly relevant to the comparison of symplectic and orthogonal predictors, since preserving geometric properties does not necessarily guarantee accurate long-term trajectories. I think experiments over several complete oscillation periods would make the conclusions more convincing.

Questions for authors
  1. How does the imposed latent symplectic structure relate to the original physical coordinates? Can the authors provide evidence that satisfying the latent geometric conditions also improves preservation of physical quantities, such as beam emittance or synchrotron tune?

  2. How are the variations in RF settings handled by the learned models? Are these settings implicitly recoverable from the input frames? If not, could the authors compare models trained under fixed machine settings or explicitly conditioned on RF parameters?

  3. Why is VISReg selected as the default regularizer when VICReg performs significantly better in Appendix C? In Table 4, VICReg reduces the mean rollout MSE ratio from 0.16 to 0.08 and the turn-249 ratio from 0.69 to 0.33. I wonder whether the main conclusions about JEPA's prediction accuracy and noise robustness would change if VICReg were used in the comparison.

  4. How general is the claimed noise robustness of JEPA? The Gaussian noise experiment shows a clear advantage, but Appendix D reports that under Poisson noise, the JEPA models experience a larger relative increase in first-turn prediction error than the autoencoder models. Could the authors clarify whether JEPA's advantage is mainly due to filtering noise in the empty background, rather than better robustness to noise affecting the actual beam distribution?

4 comments

Yes, HNN, SRNN, and SympNets partition the 64-dimensional latent vector into canonical coordinates and enforce symplecticity on the latent transition map. However, because the convolutional encoder is unconstrained, the mapping from physical phase space to latent space is not guaranteed to be a canonical transformation. Consequently, preserving symplectic volume in latent space does not automatically guarantee conservation of physical invariants, which is precisely why we observe JEPA struggling with OOD despite low latent symplectic error. Designing structure-preserving encoders that explicitly align latent variables with physical invariants or canonical variables is an open research question beyond the scope of this study, as highlighted in the Discussion section.

We thank the reviewer for highlighting this distinction. Upstream machine conditions influence the injected distribution, while the Recycler RF settings govern its subsequent evolution. In our simulations, these settings remain constant within each run, so the dynamics are autonomous conditional on the settings. However, sampling RF voltages across runs produces a family of transition maps. Our ongoing efforts toward more complex beam scenarios, including time-dependent RF programs and slip-stacking, motivate incorporating machine settings explicitly into JEPA.

Please refer to our comment at the top for our rebuttal on the rationale for choosing VISReg over VICReg.

We originally chose the 250-turn record to match the experimental oscilloscope acquisition window. We have now checked its physical informativeness using a controlled BLonD amplitude scan. These BLonD simulations save turn-by-turn coordinates for every macroparticle. A nonlinear phase-advance diagnostic using only the first 250 coordinate states recovers individual synchrotron tunes within 0.000502% of independent long-track references through 2.5 rad amplitude, resolving the amplitude-dependent frequency variation. Thus, a complete oscillation is not required for accurate frequency identification under these conditions. This supports evaluating whether JEPA captures the relevant frequencies within the recorded window.

Please refer to our comment at the top for our rebuttal on noise modeling. You are spot-on that the difference comes down to additive Gaussian noise perturbing the empty background versus Poisson noise perturbing only the occupied beam distribution. As noted above, we have narrowed our claims across the paper to specify robustness to additive noise.

AuthorsOpenReview ↗

Follow up on rebuttal

Dear reviewer,

We appreciate your time and constructive feedback on our submission. Since this is almost the last day of the discussion round, we wanted to follow up to see if our rebuttal addressed your concerns. Please let us know if we can provide any further clarification. Even if the rebuttal didn't address your concerns, we would still love to know your thoughts, since it will help us strengthen our paper.

I thank the authors for their rebuttal. I found some of the revisions in the updated manuscript help clarify my previous concerns. For example, the additional comparison between VISReg and VICReg helps explain the choice of regularizer: it seems that VICReg performs better for Linear-orth but does not work as well for the other predictor architectures, which makes the authors' choice more reasonable.

However, I think two issues still deserve further investigation, i.e., whether the imposed latent symplecticity can actually help preserve physical invariants, and how the prediction results would change if the model were explicitly conditioned on the varying RF settings.

That being said, I think many of my concerns have been reasonably touched. I will reconsider my rating accordingly.

AuthorsOpenReview ↗

Thank you!

Dear reviewer,

Thank you so much for taking the time to read our revisions and for your thoughtful follow-up. We are glad that the expanded VISReg vs. VICReg comparison and our other updates helped clarify your concerns. We appreciate your time and constructive feedback throughout the process, your engagement in the discussion, and your reconsideration of your ratings!

Reviewer_45FiOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates the usefulness of symplectic structure in JEPA world models for longitudinal beam dynamics. The authors compare several latent predictor architectures to examine whether preserving geometric properties improves long-horizon prediction. They also compare selected JEPA models with reconstruction-based autoencoders and a pixel-space surrogate in prediction accuracy and computational cost, and evaluate robustness to Gaussian and Poisson input noise and generalization to non-Gaussian initial bunch distributions. The results show that satisfying symplectic constraints alone does not guarantee better prediction accuracy, with performance depending on the predictor parameterization. JEPA trains faster and is more robust to Gaussian input noise than the evaluated autoencoders, although this advantage is not consistent under Poisson noise. Overall, the paper examines the role of symplectic structure and noise robustness in JEPA world models.

Critical strengths
  1. The paper addresses a meaningful question: whether enforcing symplectic structure in a JEPA latent predictor improves physical forecasting. By evaluating geometric properties alongside prediction accuracy, the study shows that satisfying geometric constraints does not necessarily produce more accurate rollouts. This distinction provides useful empirical guidance for designing physics-informed latent world models.

  2. The experiments provide a broad evaluation using multiple metrics, covering reconstruction and prediction accuracy, noise robustness, out-of-distribution generalization, and computational efficiency. This evaluation helps characterize the models’ strengths and limitations across several practically relevant dimensions.

Critical weaknesses
  1. The noise-robustness claim requires qualification. Gaussian and Poisson noise produce different relative performance, particularly at the first prediction step, where JEPA’s advantage under Gaussian noise does not extend to Poisson noise. The general statement that JEPA outperforms non-JEPA architectures in noise robustness needs clearer qualification by noise type and prediction horizon.

  2. The problem formulation and training objective are not presented clearly enough. The paper describes JEPA training as combining a latent-invariance term with an anti-collapse regularizer, VISReg, but does not provide an explicit equation for the complete objective. This makes it difficult to follow how the model is trained and how the loss terms are combined. A unified formulation defining each term, its weight, and the parameters it updates would improve clarity and reproducibility.

  3. The experiments lack a clear assessment of variability across training runs. The main results are presented without means and standard deviations across different training random seeds, making it difficult to assess whether the observed performance differences and model rankings are stable.

Questions for authors
  1. The Gaussian and Poisson experiments show different relative robustness between JEPA and autoencoders. The authors attribute this difference to whether the noise affects the empty background or only occupied bins. Could a controlled perturbation experiment further test this explanation? An additional noise model would also help assess whether JEPA’s robustness advantage extends beyond additive Gaussian noise.

  2. Could the authors provide an explicit equation for the complete JEPA training objective, defining each loss term, its weight, and the parameters it updates? The regularization terms are described verbally without explicit mathematical definitions. Although Appendix C.1 cites VICReg and SIGReg, their precise formulations and implementation settings should also be specified. VISReg does not appear to have a corresponding citation; please clarify whether it is an existing method or a modification introduced in this work, and provide its definition and appropriate attribution.

  3. Given VICReg’s substantially better accuracy in Table 4, the statement that VISReg uses a single fixed target “with no per-term weights” does not fully explain its selection for the main experiments. Is this choice motivated by computational efficiency? If so, could the authors compare training time and peak memory usage between VISReg and VICReg under the same setup? Could the authors also provide the Table 1 comparisons and noise-robustness results using VICReg to establish whether the main conclusions remain valid?

  4. How many independent training runs were conducted for each model, and how many different training random seeds were used? Could the authors report means and standard deviations across these runs for the main comparisons to assess whether the performance differences and model rankings are stable?

  5. The bibliography contains duplicate entries. For example, Li et al. (2021a) and Li et al. (2021b) refer to the same paper, “Fourier Neural Operator for Parametric Partial Differential Equations.” Please carefully check the reference list, remove duplicate entries, and update the corresponding in-text citations.

3 comments

Please refer to the comment at the top for our rebuttal on noise models. While a controlled perturbation experiment can characterize this specific convolutional encoder, the behavior will still depend heavily on the chosen encoder architecture and cannot be generalized to all JEPA networks. We have therefore revised our claims throughout the paper to specify the robustness of our convolutional JEPA implementation to additive noise.

We have added the complete training objective as Eq. (2) in Section 5 alongside a detailed breakdown in Appendix C.1 defining every loss term, its weight, and the exact parameters each term updates.

Please refer to the comment at the top for our rebuttal and rationale for selecting VISReg over VICReg. At D=64, per-epoch wall-clock training time (4.63 vs. 4.52 min/epoch) and rollout inference time (10.06 vs. 9.11 s) are nearly identical (reported in Appendix C.1 Table 5).

Each configuration in our benchmark grid was trained once with a fixed seed because training across the full full-grid phase-space corpus is computationally expensive. However, we evaluate every trained model across multiple held-out test trajectories over 249-turn autoregressive rollouts and across multi-seed noise draws, where model rankings and failure modes remain highly consistent across horizons.

We thoroughly reviewed the paper, removed duplicate entries, and updated the corresponding in-text citations.

The revised noise-robustness claims, the explicit training objective, and the expanded comparison between VISReg and VICReg address most of my concerns.

AuthorsOpenReview ↗

Thank you!

Dear reviewer,

We appreciate your response to our rebuttal!

We are glad to hear that the expanded VISReg vs. VICReg experiments, explicit training objective, and refined noise-robustness claims addressed most of your concerns. If you have any remaining questions or concerns, please feel free to share them, and we would be happy to provide further clarification. Our study, like any other, is not perfect, but it provides rigorous evidence on how physics-informed JEPAs behave on complex, real-world Hamiltonian dynamics and offers practical guidance for future world-model design. We truly appreciate you updating your evaluation after our initial response. Since most concerns from your initial review have now been resolved, we would appreciate it if you could revisit your scores and weigh them against this remaining scope question.

That said, as we mentioned earlier, we appreciate your time and effort in reviewing our paper, engaging with us during the rebuttal, and helping us strengthen it through constructive feedback.

P.S. We also noticed that "Flag For Ethics Review: Yes" is currently checked on your review form. Since no ethical concerns were mentioned in the review, could you please check if that was selected by mistake?

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

The authors aim to determine whether JEPA architecture can learn dynamics of a Hamiltonian system to investigate whether JEPA is a good architecture for learning complex physical dynamics.

In short, the paper aims to answer the following questions: 1/ when one trains JEPA architecture on sympletic system, does JEPA model restore the geometry of the one-turn map, and improve accuracy 2/ how JEPA compares with reconstruction-based and neural-operator baselines 3/ whether JEPA is robust to distribution shift and observation noise

To check this, the authors replace the predictor of the baseline JEPA (which uses MLP as predictor) by a number of "structure preserving" predictors (HNN, etc), and see how does these JEPA-encoded predictors do in comparison to each other and other baselines.

Critical strengths

The paper has some critical strengths:

1/ It was an interesting test case where one can measure beyond simple forecasting errors (like RMSE) 2/ the comparisons are well controlled (or so it seems) 3/ The paper examines multiple noise models

Critical weaknesses

1/ The paper doesn't define certain terms very often (and missing citations of relevant work) (eg. VISReg) 2/ It is hard to draw any generalizable conclusions from testing JEPA on a simple hamiltonian system to say more complex dynamical system. Most real life systems are not phase space preserving hamiltonian systems at all. 3/ The paper’s own ablation studies undercuts its conclusion. The main comparison (Table 1) uses the VISReg regulariser and concludes that autoencoder gives the lowest error and that JEPA doesn’t track the bunch centroid. However, in Table 4, Linear-orth (JEPA) trained with VICReg nearly matches the autoencoder on every long-horizon metric. So it is confusing to say the least.

Questions for authors

Please clarify at least the following:

The main comparison (Table 1) uses the VISReg regulariser and concludes that autoencoder gives the lowest error and that JEPA doesn’t track the bunch centroid. However, in Table 4, Linear-orth (JEPA) trained with VICReg nearly matches the autoencoder on every long-horizon metric. So it is confusing to say the least. Would you mind commenting on this?

3 comments

Please refer to our comment at the top for our rebuttal on the rationale for choosing VISReg over VICReg.

Regarding your observation on the centroid claim: we explicitly note in Sections 6 and 7 that while JEPA models under VISReg match the persistence baseline on centroid error, pairing Linear-orth with VICReg recovers the autoencoder's centroid accuracy (0.006).

Lastly, we added a citation for VISReg and made a thorough pass to ensure all terms are well-defined.

Reviewer_zxNWOpenReview ↗

Brief response to your comment

Dear Authors,

I thank you for clarifying the definitions and various choices such as VISReg vs VICReg ,it helps clarify a lot of the issues in the paper.

However, as I mentioned in the original review, it is important to show some kind of generalization before this one system , even though it is a real world system before making a seemingly broad statement (that really so far affects one system that is being investigated).

AuthorsOpenReview ↗

Thank you!

Dear reviewer,

We appreciate your response to our rebuttal!

We completely understand your perspective regarding testing across multiple systems. While we could have included simpler synthetic benchmarks (such as a pendulum or mass-spring system), prior work has already explored JEPAs on low-dimensional pendulums, and those toy systems do not capture the complexity of a particle physics accelerator. Curating 23,101 high-fidelity BLonD simulations of the Fermilab Recycler required deep domain collaboration with particle physics experts to establish a realistic benchmark, and building an equivalent high-dimensional corpus for a second complex physical facility does not scale trivially within a single study. However, if you have specific concerns about the properties of our dataset that make you doubt the generalizability of our findings (for example, regarding how the Hamiltonian structure manifests in the data or uncertainty quantification), please let us know, and we would be glad to provide additional data diagnostics or statistical analyses to address them. While we fully respect your judgment, we would really appreciate it if you could consider revisiting your scores and weighing them against this remaining scope question.

That said, as we mentioned earlier, we appreciate your time and effort in reviewing our paper, engaging with us during the rebuttal, and helping us strengthen it through constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.