acceptodds
Under review as a conference paper at ICLR 2027

Identify then Realize: Contrastive Learning of Latent Port-Hamiltonian Dynamics from Partial Observations

34 views

Abstract

Identifying latent state representations and dynamics is essential when direct modeling in observation space is infeasible, particularly under partial and high-dimensional observations. In such settings, representation learning and physics-aware modeling are inherently coupled. We propose CIPHER, a two-stage identify-then-realize framework for learning latent port-Hamiltonian models of conservative and dissipative systems. First, a contrastive teacher jointly learns an encoder and a neural ODE to obtain predictive state representations from observation histories. Second, a student learns an invertible nonlinear coordinate transformation and port-Hamiltonian dynamics by matching encoded observed futures. We establish sufficient conditions for recovering the physical state up to a diffeomorphism and realizing its dynamics in port-Hamiltonian form. Across ten clean and noisy datasets, CIPHER is competitive with or outperforms state-of-the-art baselines by improving long-horizon forecast on complex systems. Further studies show robustness to over-specified latent dimensions and additional predictive gains from pH parameterization.

Primary area:
learning on time series and dynamical systems
Keywords:
Physics-informed Machine Learning, Contrastive Learning, System Identification

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_ZAT8OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces CIPHER, a two-stage method for learning physical dynamics from partial or noisy observations such as positions or images. First, a contrastive encoder and neural ODE learn predictive latent states from observation histories. Then, a second model learns an invertible coordinate transformation and port-Hamiltonian dynamics that uses the frozen representations. The paper provides conditional theoretical guarantees and reports forecasting improvements on synthetic conservative and dissipative systems.

Critical strengths

1- The paper addresses a meaningful identification problem. Learning structured dynamics from partial observations is a good topic to focus because positions or images often omit essential state variables such as momentum.

2- The two-stage framework is motivated because separating representation learning from structured dynamics fitting simplifies the student’s optimization by providing fixed latent targets.

3- Under explicit observability and information-retention assumptions, the paper establishes smooth state recovery and a bridge to compatible pH dynamics (although limit the systems to current pH setting, but still worth to have it).

4- The reported results demonstrate useful forecasting improvements on complex systems. CIPHER improves over its teacher on the three-mass chain and clean and noisy three-body conditions at the reported 50-step horizon.

5- The ablations exps and reporting help assess the design choices to involve the student.

Critical weaknesses

1- The pH assumption in its current form for physical latent is somehow restrictive because it excludes physical behaviors in the main same-dimensional setting as your assumption A3 supposes a compatible pH realization. Considering what you have in your assumptions for smoothness, unforced model with R≥0R \geq 0 cannot be expressive. Please consider this for example: an attracting periodic orbit on a full neighborhood in the same state space. This is a scope restriction, although broader pH models with inputs or additional states can manage to fix (and such case is not in your exps). Please write this clearly as you limitation and scope of the work.

2- The current evaluation (as much I can see in main and appendix) has limited evidence of reliability over longer horizons. In other words, I specifically mean training includes forecasts up to 32 time steps, while testing extends to 50. Small prediction errors may accumulate over hundreds of steps, causing artificial damping, energy drift, or unstable trajectories (which is not feasible to see now because of short horizon). Longer rollouts, with measurements of energy conservation and dissipation could help establish the effectiveness of your pH structure. I would suggest the authors to include this longer horizon testing for their benchmarks already studied in paper (e.g. rollouts with horizons 100-200-500).

3-The temporal cutoffs used for negative sampling (in main body they are l and r) are important HPs that suitable values may depend on the system’s setting. Although in your appendix I can see specifies (l=5) and (r=20), there is no argument about selection or evaluate sensitivity to them. Unsuitable cutoffs could retain recurrent states as negatives or exclude useful comparisons, potentially degrading representation learning (if not saying ruin the learning by increasing false positive and false negative in your set). A sensitivity analysis across the existing benchmarks, together with guidance for selecting these cutoffs on new systems is important for me to see how scalable your approach is (I am of course not asking to provide huge reimplementation of your experiment. A simple additional report is enough though, that I believe with your current code repo is feasible to achieve).

Questions for authors

I wonder to know about some extensions

1- Current setting is targeting a deterministic framework, which is fine but could be extended to probabilistic inference rather than hardcoded loss. Could you explain how such achievement is feasible ? (this is an important extension btw, because you will be able to quantify the uncertainty of your predictions at the end rather than sticking with 100% sure answers).

2- The sensitivity analysis for physical state size is good addition that you have but I am assuming you are aware of physical state size, which is relatively low dimensional. I wonder to know what if we consider a system with unknown and higher dimensional physical latent and then see how your model works (a simple example would be a traffic scene with several cars, then for each car physical latent would be 3 size position and 3 size velocity (ignoring acceleration) times by number of cars). How do you think your setting extend to such case? This can extend the scope of your work as well, if answered appropriately.

In the current setting I will give a borderline score to your paper (which is not listed in scores), but would be happy to reconsider depending on how you address the concerns and questions.

4 comments
AuthorsOpenReview ↗

Rebuttal addressed for reviewer ZAT8 (1/2)

Thank you for the thoughtful and constructive review. We appreciate your recognition of the staged identification–realization framework and the theoretical connection between state recovery from partial observations and structured dynamics modeling. Your suggestions have helped us sharpen the CIPHER’s scope and strengthen its evaluation. We discuss the extension questions in a separate comment that follows.

W1: Scope of pH realizability. We thank the reviewer for highlighting this limitation. We agree with you that our smooth, unforced pH student cannot represent an attracting non-equilibrium periodic orbit on a full open neighborhood counterexample you mentioned, and a diffeomorphic coordinate transformation cannot remove this obstruction.

A3 is therefore a substantive compatibility assumption. Theorem 1 separates state recovery under A1–A2 from pH realization under the additional A3, and it does not claim universal pH realizability. The compatible class nevertheless includes practically important systems like free vibration of buildings/bridges, protein conformational relaxation, and so on, in addition to systems already in experiments (Togashi and Mikhailov, 2007).

Input-driven and augmented-state formulations offer routes to broaden this scope. Explicit input ports can represent external energy supply sustaining oscillations, while suitable state augmentation may permit a higher-dimensional structured realization whose projection reproduces the target dynamics. These require additional modeling assumptions. We clarified this restriction after A3 and discuss input-driven and augmented-state extensions with supporting references.

W2: Longer-horizon evaluation. Following your suggestions, we updated Table 1 which now extends evaluation to 100 steps. We also strengthen the comparisons as reviewer q2jn suggested: Mamba-3 with matched observed frames and supervised horizons, and PHAST is additionally trained with supervised horizon loss.

At 50 steps, CIPHER still achieves the highest mean (R^2) in all four noisy three-body conditions. At 100 steps, its advantage extends to all six three-body conditions, including clean dynamics, with the highest reported mean (R^2@100) and (AUC1:100)(\mathrm{AUC}_{1:100}). In particular, although rollout-trained PHAST leads at the clean three-body comparisons at 50 steps, CIPHER retains substantially greater accuracy at 100 steps: 0.770 versus 0.299 on planar dynamics, and 0.764 versus 0.078 on 3-D dynamics. Overall, for complex dynamics tasks, CIPHER performs substantially better than all other methods for longer horizon performance.

Furthermore, CIPHER achieves the longest mean horizon before (R2≤0)(R^2\leq0) in all six clean/noisy three-body conditions. On clean planar and 3-D dynamics, it reaches (301±33)(301\pm33) and (345±51)(345\pm51) steps, compared with (23±2)(23\pm2) and (16±2)(16\pm2) for matched Mamba-3, and (111±5)(111\pm5) and (103±2)(103\pm2) for rollout-trained PHAST. Across all six conditions, CIPHER’s mean horizon is 1.37–3.35 times that of the strongest external baseline in each condition. These results across seeds demonstrate that CIPHER supports substantially longer autonomous forecasts, extending its advantage well beyond the 100-step benchmark.

W3: Negative-band sensitivity. We thank the reviewer for suggesting this analysis, now included in Table 3. CIPHER’s identification stage is generally insensitive to the tested cutoff variations, with two noticeable exceptions. We compared four alternative bands against the default ((5,20]) on pendulum, noisy three-body ((\sigma=0.01)), and three-mass dynamics, using three seeds per setting. Across these twelve comparisons, ten change mean (R^2@50) by at most 0.4 percentage points. Because (l) and (r) directly affect contrastive teacher training, we use the paper’s teacher-rollout protocol.

Pendulum remains near ceiling consistently, and varying the upper cutoff from 10 to 40 has little effect across all tasks. Both exceptions involve the lower cutoff: decreasing (l) to 2 reduces the chain’s score from 0.975 to 0.950, while increasing (l) to 10 reduces the three-body score from 0.886 to 0.839. These results support broad stability within the tested range while identifying some task dependence in the lower cutoff.

Rationale and guidance for new systems. Despite observed robustness, we provide practical guidance for applying CIPHER to new systems. For a positive target at (t+h), same-trajectory negative candidates are retained when (l<|j-(t+h)|\le r). The lower cutoff excludes nearby, potentially redundant states; the upper cutoff limits distant candidates that may revisit similar states. We recommend expressing the cutoffs in physical time, (l\Delta t) and (r\Delta t), to account for sampling-rate differences. Slower local variation may motivate a larger (l), while shorter recurrence times may motivate a smaller (r).

AuthorsOpenReview ↗

Rebuttal addressed for reviewer ZAT8 (2/2): Extension questions

Thank you for the extension questions. CIPHER’s separation of state identification and dynamics realization provides a useful platform for both directions.

Q1: Probabilistic extensions. We agree that uncertainty quantification is an important extension. CIPHER’s separation of representation learning and dynamics realization provides a natural interface for probabilistic modeling. One concrete direction is a Gaussian process prior on the student Hamiltonian model, following from Beckers et al, (2023) and Ewering et al, (2026), while retaining skew-symmetric (J) and positive-semidefinite (R) with their parameters estimated.

Conditional on the learned encoder and invertible chart, the estimated latent trajectories would provide data for probabilistic dynamics identification. Predictive trajectories could then be generated by sampling posterior Hamiltonian functions, integrating their corresponding pH fields, and mapping the trajectories back to observation space. This would quantify dynamics uncertainty conditional on the learned representation.

A complementary extension is a probabilistic history encoder, allowing uncertainty about the initial state to propagate through the student. Combining these components would address both representation ambiguity and dynamics uncertainty. Estimated latent states are not exact physical-state measurements, so an appropriate likelihood and uncertainty treatment would be needed. Calibration should be assessed in observable quantities using predictive coverage and proper scoring rules. These extensions preserve the staged modeling principle but require additional inference machinery and validation.

Q2: Traffic modeling. Traffic raises two related questions: selecting latent capacity when the physical state dimension is unknown, and scaling the realization to many interacting agents.

Unknown state dimension. CIPHER treats latent dimension as a modeling choice. It does not require that the selected dimension equals the minimal physical-state dimension, and our dimension study (Figure 3 and Table 9 in the paper) shows tolerance to over-specification on the evaluated systems. For N vehicles, positions and velocities provide an initial kinematic estimate—for example, 6N coordinates in three dimensions—but road constraints and unobserved behavioral variables may increase the information needed for predictive state. We therefore view such counts as starting points for capacity selection, not as known ground-truth dimensions.

Vehicle-level particle model and composition. A promising extension would infer vehicle-level states from observation histories, then learn controlled pH subsystems and their interactions. Vehicle mechanics provide natural components: kinetic energy, dissipative forces, and traction/braking inputs. There is also direct precedent for pH traffic modeling: Ackermann et al. (2025) formulate stochastic, controlled car-following dynamics using an input-state-output pH model, studied with a fixed vehicle population and periodic boundaries.

For composition of these vehicle subsystems, Neary and Topcu (2023) develop a framework that learns pH neural models of individual subsystems and combines them through known or learned interconnection structures. This suggests a modular extension of CIPHER: identify predictive representations first, freeze them, and then fit local structured dynamics and compatible couplings. Shared vehicle models and sparse interactions could improve scalability compared with a single dense model of the entire traffic scene.

A concrete extension path. A practical setting is controlled car-following with a fixed vehicle population: identify vehicle-level predictive states, learn local pH dynamics with compatible input/output ports, and connect them through structure-preserving interactions. Driver actions and external influences can be incorporated through input or behavior models. Variable-agent representations and hybrid transitions would then provide a route to handling arrivals, departures, and lane changes. This progression builds directly on CIPHER’s identify-then-realize design, providing a concrete path from the present autonomous formulation toward controlled, interacting traffic systems.

Reviewer_ZAT8OpenReview ↗

Rebuttal response

Thanks for the rebuttal and answering my queries. I'm satisfied. Well done!

In case of acceptance please make sure to put the future directions and stochastic variant in discussion/future works part of your paper.

I originally gave borderline grade, after rebuttal I will increase it to accept. Although it is a good work, but still there is plenty of rooms to improve the work through the extensions that authors talked about, and thus I dont consider max grade (4) for this current version. However my evaluation score is above weak accept as well. Since there is not score between weak accept (3) and max grade (4), I put score 4.

AuthorsOpenReview ↗

Response to reviewer ZAT8

Thank you for carefully considering our rebuttal. We sincerely appreciate your constructive feedbacks throughout the review process. As you suggested, we will clearly discuss the proposed extensions and directions in the paper, and we plan to pursue them in near future work.

Reviewer_q2jnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The submission addresses the issue of learning the dynamics of mechanical systems when the state is only partially observed, such as when there are an angle without its velocity information or a video frame is missing. The paper proposes CIPHER, which is a 2-stage network, where the first stage trains a teacher that uses a neural ODE backbone to learn future latent representations. Stage 2 further freezes the encoder and trains a student model as well as an invertible map into new coordinates. On the three-mass chain and three-body systems the paper demonstrates that CIPHER obtains the best mean long horizon accuracy.

Critical strengths

The key insight the paper provides is a concrete reason why particular physics constraints (hamiltonian dynamics) is difficult to impose on latent representations that are partially observed (missing frame etc). If a netywork attempts to train both the encode and latent dynamics together the field becomes constantly changing target that neither is able to well approximate. The proposed method freezes the encoder and them fits the coordinate change and field, removing the moving target conundrum. Both the one-stage and two-stage training share the underlying objectives and backbones so comparison to isolate the gain is simple.

Critical weaknesses

W1. The port-Hamiltonian formualtion feels a little cosmetic (sorry...), which kinda undercuts the central selling point of the submission. The second stage replaced with a unconstrained neural ODE student is able to recover most of the performance compared to the version with structural hard constraints. No energy or dissipation graphs to tables are used as diagnositics to show that the port-Hamiltonian formulation is doing real regularization work on the learned latent space.

W2. The baselines feels artificially gimped. Mamba-3 sees only 10 frames whereas the teacher observes 21, which makes it difficult to isolate where the performance delta relative to the baseline is coming from. Furthermore DCL uses a single linear latent transition and I struggle to see how the network setup is able to well apprxoimate a 3-body system. Lastly PHAST trains on teacher forcing with no rollout loss which the proposed methods received 32 horizon. I think the authors need to revisit how they set up their experimental evaluation. Before such reflection happens it is hard to call the proposed method SOTA when it is compared to baselines on clearly unequal footing.

Questions for authors

Please address the two question in the weakness section.

4 comments
AuthorsOpenReview ↗

Rebuttal addressed for reviewer q2jn (1/2): Contribution and role of the pH formulation.

Thank you for the thoughtful and constructive review. We appreciated your comments and your recognition of CIPHER’s staged design for learning structured dynamics. Your comments have helped us to clarify the role of pH structure and strengthen the comparisons. Please see our rebuttal to your comments:

W1: Contribution and role of the pH formulation.

Additional predictive benefits. In Table 7, CIPHER improves mean (R^2@50) over NODE from 0.859 to 0.877 on clean 3-D dynamics and from 0.772 to 0.811 on planar dynamics at (\sigma=0.05): 12.8% and 17.1% lower normalized prediction error (1-R^2). At 100 steps, all five seed pairs favor pH in each of the three 3-D conditions (Table 8). Table 9 shows positive mean paired relative reductions at every tested dimension above the physical state size on the chain and both three-body systems. Both students implement our identify-then-realize framework. NODE's strong performance supports the shared second stage, while these ablations establish additional pH benefits.

Meaningful energy and dissipation. CIPHER provides an explicit (H) and enforces (H˙=−∇H⊤R∇H≤0)(\dot H=-\nabla H^\top R\nabla H\leq0) in continuous time. Our new diagnostics test correspondence with true energy. Appendix Fig. 7 shows CIPHER learned Hamiltonian visually match with pendulum energy contours and Duffing's double-well geometry. We also measure learned and true energy during 50-step forecasts using Spearman correlation, and compare their average normalized energy-change rates on held-out observed states, reporting learned rates across five seeds:

SystemEnergy (\rho), meanMedianNorm. learned rateNorm. physical rate
Pendulum video0.9960.997−0.004 ± 0.0020
Duffing video0.9360.939−0.247 ± 0.015−0.244
Three-mass chain0.8810.940−0.145 ± 0.007−0.135

CIPHER learns meaningful energy representation and dissipation without ground-truth energy or dissipation supervision in training. Mean energy correlations reach 0.996, 0.936, and 0.881 for pendulum, Duffing, and the three-mass chain, respectively. Average normalized dissipation magnitudes agree with their physical counterparts within 1.2% for Duffing and 7.4% for the chain, while remaining near perfect for pendulum. Thus, correspondence with true energy emerges from learning the dynamics under the pH field. The unconstrained student has no energy function, so no such diagnostic exists for it. Together with the student stage ablations, they provide additional evidence of meaningful learned energy and additional predictive benefits.

In addition, this explicit energy representation in PHS opens doors for powerful tools for analyzing and controlling systems. In fact, many electromechanical systems can be expressed in Port-Hamiltonian form (van der Schaft and Jeltsema, 2014). Literature includes powerful tools that build upon PHS model: energy shaping supports stabilization under appropriate matching and stability conditions (Ortega et al., 2002), compatible interconnections enable compositional modeling (Neary and Topcu, 2023), and so on. These capabilities motivate an explicit pH realization, which CIPHER extends.

We address the requested stronger baseline controls and updated comparisons in the next comment, “Rebuttal addressed for reviewer q2jn (2/2): Strengthened baseline comparisons.”

Together, the controlled student comparisons and new diagnostics demonstrate additional predictive benefits, meaningful energy representation, and aggregate dissipation matched with physical behavior—all without ground-truth energy or dissipation supervision. The pH formulation therefore contributes both forecasting benefits and an explicit structured representation that supports further analysis and control.

AuthorsOpenReview ↗

Rebuttal addressed for reviewer q2jn (2/2): Strengthened baseline comparisons.

This comment addresses W2; our companion response (1/2) addresses contribution and role of the pH formulation.

W2: Baseline comparisons. We thank the reviewer for identifying and suggesting stronger experimental controls. We matched the observation context and supervised horizons for baselines and extended evaluation to 100 steps—more than three times the maximum training horizon of 32 steps (Table1).

  1. Matched Mamba-3. CIPHER achieves higher (R^2) on all seven higher-dimensional, complex tasks at both 50 and 100 steps than improved Mamba-3. Our original baseline configuration used a default context of 10 observed frames. We agree with you and have retrained Mamba-3 with 21 observed frames and supervision through 32 steps. Increasing context improves some Mamba-3 forecasts, but CIPHER retains higher mean (R^2) on the three-mass chain and all six clean/noisy three-body conditions at both evaluation horizons. For example, on most noisy 3-D three-body dynamics, CIPHER achieves 0.807 versus 0.627 at 50 steps, and 0.495 versus 0.037 at 100 steps.

  2. Matched PHAST. CIPHER leads on all six clean/noisy three-body tasks in long horizon performance. Our original choice followed the training objective used in PHAST’s published position-only forecasting experiments PHAST paper. Our choice of PHAST was intended to follow the published experimental protocol.

    We nevertheless agree with you and trained matched PHAST with 21 observed frames and rollout supervision through 32 steps, matching CIPHER’s training. Improved PHAST model now leads on the three-mass chain and at 50 steps on clean three-body dynamics. However, CIPHER retains its advantage under noisy observations and achieves substantially stronger long-horizon forecasts on the three-body systems. On most noisy 3-D dynamics, CIPHER achieves 0.807 versus 0.667 at 50 steps, and 0.495 versus 0.071 at 100 steps. On clean planar/3-D dynamics, CIPHER achieves 0.770/0.764 at 100 steps, versus 0.299/0.078 for matched PHAST.

  3. DCL. We acknowledge the limitation of single-linear-transition. Its switching rule selects a mode using the next encoded observation (Eq. 6), unavailable during autonomous forecasting. This extension would require a new causal mode selector. We retain the three documented input variants and explicitly included these limitations for forecasting already in the appendix B.2. Furthermore, our claims that CIPHER is closely competitive or outperforms other baselines do not rely on DCL as we recognized DCL’s limitation for forecasting.

Overall, CIPHER’s advantages persist after addressing the identified context and training-horizon differences. Its pH student adds predictive benefits and learns meaningful energy–dissipation structure without ground-truth energy supervision.

Thank you for the in depth rebuttal. I think these do ultimately address my concerns. Comparing the manuscript with the old version I also see that the authors made efforts to incorporate these new experiments and discussions. I have raised my score to 3. Best of luck.

AuthorsOpenReview ↗

Response to reviewer q2jn

Thank you for considering our rebuttal. We sincerely appreciate your constructive feedbacks through the review process, and we are glad that our rebuttal addressed your concerns.

Reviewer_dV9eOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

CIPHER learns a predictive history representation with a contrastive encoder and neural-ODE teacher. It then freezes that representation, and fits a port-Hamiltonian student to encoded observed futures. An invertible chart me- diates the structured realization. The intended benefit is to separate representation learning from imposing physical structure when observations omit state variables.

Critical strengths

Th observed-future versus teacher-rollout targets test error inheritance. Further, an unconstrained second-stage student controls the task of refitting. Next coordinate-by-matrix comparisons distinguish the role of a learned chart from state-dependence. A shared frozen forecasting probe and a parameter-matched latent-dimension comparison help to improve interpretability.

Critical weaknesses

The incremental value of the structured field is modest and insufficiently quantified. Against the unconstrained refit, the pH mean advantage at 50 steps is only 0.002-0.036 across the chain and six three-body conditions, without paired uncertainty.

The evidence only appears to support restricted forecasting. All conditions are simulated. The largest supervised horizon is 32 steps and evaluation reaches 50. Further, external baselines differ in capacity, con- text, and objectives.

Questions for authors
  1. What are the paired seed-level differences between structured and unconstrained students, and which structural gains remain reliable after accounting for variability?
  2. Which pipeline or coordinate findings persist under a comparably expressive one-stage alternative with matched training resources?
1 comment
AuthorsOpenReview ↗

Rebuttal addressed for reviewer dV9e

Thank you for the thoughtful review and for your recognition of CIPHER’s staged design for learning structured dynamics. Please see our rebuttals to your comments on paired-seed comparison and baselines evaluation:

W1/Q1: Paired-seed difference between PHS and Neural ODE students. Both NODE and pH students instantiate our identify-then-realize framework: they learn dynamics from frozen predictive representations using encoded observed futures. We agree that much of the improvement is attributable to second-stage learning, which is shared by the unconstrained NODE and pH students. To quantify the additional pH benefit, we now report paired-seed differences across five seeds, including their sample standard deviations in Table 8 of the paper. For convenience, we provide a short overview in the following:

The additional pH benefit persists at longer horizons and becomes larger in several challenging settings. At 50 steps, the mean paired difference favors CIPHER on all ten tasks. At 100 steps, CIPHER improves mean dynamical (R^2) by 0.0232–0.0274 across all three clean/noisy 3-D three-body settings, with positive differences in all five seed pairs in each setting. At this horizon, every pair also favors CIPHER on the three-mass chain and planar three-body dynamics at (\sigma=0.05). The mean advantage increases from 0.0183 to 0.0274 on clean 3-D dynamics and from 0.0389 to 0.0491 on noisy planar dynamics when extending evaluation from 50 to 100 steps. Thus, the additional pH benefit on these complex systems is consistent across seeds and persists beyond the original evaluation horizon.

In addition to the quantitative improvements, the rich structure of the port-Hamiltonian student further opens the door for control theoretic approaches such as energy-shaping methods, reachability and stability analysis, system interconnection, and other energy-based tools((van der Schaft and Jeltsema, 2014; Ortega et al., 2002; Neary and Topcu, 2023). These capabilities motivated our choice of using PHS models, which CIPHER extends.

W2: Extended baseline comparisons. We extended the evaluation horizon to 100 steps (more than 3x longer than training horizon) for all methods, and we matched objective and context to baselines whenever applicable. Please refer to Table 1 of the updated paper. Rollout-trained PHAST leads by a margin on the three-mass chain and at 50 steps on the clean three-body tasks. CIPHER achieves the highest reported mean (R^2@50) on all four noisy three-body conditions, and the highest mean (R^2@100) and (\mathrm{AUC}_{1:100}) on all six clean/noisy three-body conditions. To highlight CIPHER’s performance in long horizon, for clean planar/3-D dynamics, its (R^2@100) is 0.770/0.764 versus 0.299/0.078 for rollout-trained PHAST. In other words, CIPHER achieved 3.05x/3.91x lower normalized prediction error than PHAST in those tasks. These stronger comparisons support CIPHER’s forecasting advantages under noisy observations and at longer horizons for complex dynamical systems.

For our experiment setting, we believe context and objectives are most important for forecasting. From our baselines, Mamba-3 would be most flexible model, while the results indicates even with matched context and objective, yet CIPHER outperforms strengthened Mamba in harder tasks. However, we would welcome any suggestions for additional baselines to include.

Simulation and observation noise. Controlled simulation is an established evaluation approach in structured dynamics learning (Greydanus et al., 2019; Zhong et al., 2020). Our 10 datasets span conservative and dissipative systems, numerical and visual observations, low and high dimensional system, clean and noisy observations, and simple and well-known complex dynamics. For both planar and 3-D three-body systems, the higher per-coordinate noise scale, 0.05, is comparable to the median planetary displacement between frames, making inference of dynamics from noisy positions substantially more challenging. While evaluation on real data remain future work, CIPHER has demonstrated strong performance across these comprehensive experiments.

Q2: Capacity-matched one-stage versus two-stage learning. Following your suggestion, we enlarged the one-stage constant JR pH model to approximately match the combined teacher-and-student models parameter size and training resource. Therefore, they are comparably expressive. Although enlarging the baseline improves its performance, two-stage CIPHER retains higher mean (R^2) on eight of ten tasks at 50 steps and seven at 100 steps, including all three 3-D and both noisy planar conditions (see Table 6). At 100 steps, it achieves 0.678 versus 0.530 on clean 3-D dynamics and 0.419 versus 0.110 at (\sigma=0.05). These results further show that the staged procedure’s advantages, especially on complex and noisy systems, persist after the additional capacity for one-stage pipeline.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We propose CIPHER, a two-stage identify-then-realize framework for learning latent port-Hamiltonian models of conservative and dissipative systems. First, a contrastive teacher jointly learns an encoder and a neural ODE to obtain predictive state representations from observation histories. Second, a student learns an invertible nonlinear coordinate transformation and port-Hamiltonian dynamics by matching encoded observed futures.

We sincerely thank all reviewers for their thoughtful and constructive feedbacks. Building on the strengths already recognized by the reviewers, our rebuttal further strengthens CIPHER’s empirical evaluation and clarifies the scope of its theoretical claims. The main changes are:

  • Stronger baselines and substantially longer forecasts. Reviewers questioned differences in baseline training and the original evaluation horizon. We retrained Mamba-3 and PHAST with matched observation context and forecast supervision, then extended evaluation to 100 steps (more than 3x longer than training horizon) and beyond. CIPHER’s forecasting advantage persists under these stronger comparisons: it achieves the best mean long-horizon scores across all six high-dimensional and complex three-body conditions with both clean and noisy data and sustains predictive performance substantially longer than the external baselines.

  • Clearer evidence for staged learning and meaningful pH structure. Reviewers asked whether the gains mainly came from second-stage learning and whether the pH formulation added substantive value. Our added paired-seed comparisons and enlarged one-stage pipeline study justifies CIPHER’s identify-then-realize approach and supports additional predictive benefits from pH structure. Crucially, new energy and dissipation diagnostics show that CIPHER learns meaningful energy representations and dissipative behavior without ground-truth energy or dissipation supervision.

  • A more precise theoretical scope. We explicitly acknowledge the reviewer’s counterexample and distinguish state recovery from the additional assumptions required for pH realization. The revised theory clearly states this limitation. We also outline cited input-driven and augmented-state extensions, while keeping them separate from the current guarantees.

  • Evidence of robustness and practical guidance. The requested negative-sampling study shows that identification-stage forecasting is largely stable across the tested cutoff variations. We further provide guidance for selecting these settings on new systems. Separately, we outline concrete routes toward probabilistic dynamics modeling and compositional traffic modeling, with both extensions naturally building on CIPHER’s separation of representation learning and structured realization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.