- 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 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
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.
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.
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.
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). 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) in all six clean/noisy three-body conditions. On clean planar and 3-D dynamics, it reaches (301±33) and (345±51) steps, compared with (23±2) and (16±2) for matched Mamba-3, and (111±5) and (103±2) 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).