acceptodds
Under review as a conference paper at ICLR 2027

Stable and Counterfactually Robust Physical World Models from Imposed Structure and Learned Physics

Abstract

A world model learns to forecast how a physical system evolves from recorded trajectories, yet the systems it imitates obey physical laws that are neither fully supplied nor reliably respected. The model may create energy, drift or diverge over long rollouts, and answer a changed-law query using the law observed during training. We ask how much general physical structure must be hard-coded into a world model, and how much system-specific physics can then be learned from data, for four properties to hold simultaneously: second-law-compatible dissipation, correct responses to interventions on physical parameters, stability out to one hundred times the training horizon, and robustness to disturbances. The imposed structure is general: dynamics are generated from the gradient of a learned energy through a fixed reversible operator, the energy is restricted to a confining class, a one-way port can remove energy but never inject it, the drive channel is known, and the intervened parameter enters through a separable map. The model learns the energy functional, constitutive relations, dissipation rate, and couplings. Across an electromagnetic cavity, a particle-in-cell grid, and a shallow-water fluid, models with roughly nine thousand parameters recover constitutive functions with unit slope, separate conserving from dissipating worlds by four orders of magnitude using a single set of weights, and transfer changes in sign, magnitude, rate, and gravity to unseen values, where equal-capacity models without the same structure perform at chance or worse. A nonlinear constitutive law is recovered with its curvature preserved and predicts a held-out intervention – better than a converged linear model. The same structured model also refuses interventions for which no parameter map was supplied, producing errors to larger than in-distribution, while oracle models given the missing map succeed; its one-way port likewise cannot represent anti-dissipation by construction.

Primary area:
foundation or frontier models, including LLMs
Keywords:
World Model, Physics discovery, Long term stability, Counterfactual robustness

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_eQRoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a model that incorporates general physical structure into neural networks while learning system-specific physics from data. The imposed structure includes energy conservation, the form of dissipation, and how physical parameters such as charge and gravity enter the governing equations. The learned components include the energy function, dissipation rates, and constitutive relations. The authors evaluate their approach on three physical systems: an electromagnetic cavity, a particle-field simulation, and a shallow-water fluid. They experimentally demonstrate that the model can distinguish conservative from dissipative systems, identify nonlinear constitutive relations, extrapolate to unseen physical parameter values such as reversed charge and modified gravity using prescribed parameter dependencies, and maintain long-term stability over horizons up to 100 times the training horizon. The paper claims four main contributions. First, building on structure-preserving approaches such as Hamiltonian Neural Networks (HNNs), it integrates physical constraints on energy dissipation, extrapolation to unseen physical parameters, long-term stability, and robustness to disturbances within a unified framework. Second, it focuses on interventions on physical parameters (counterfactuals) and establishes conditions under which the model can correctly predict the resulting dynamics. Third, through ablation experiments that remove individual structural components, it investigates which components are responsible for each desirable property. For example, the authors show that long-term stability depends not only on the properties of the energy function but also on the consistency of energy exchange between particles and fields. Fourth, the paper introduces a mechanism for selecting among predefined candidate physical model structures using the Bayesian Information Criterion (BIC). It further applies conformal prediction to provide a finite-sample guarantee, under certain statistical assumptions, on the probability of retaining the true model within the candidate set.

Critical strengths

One of the major strengths of this paper is that it presents a clear research concept: incorporating known physical structures and parameter dependencies into a model while learning unknown constitutive relations and dissipative properties from data, with the aim of enabling extrapolation to physical conditions not observed during training. In particular, I believe that this research direction, which combines physical constraints with data-driven model inference to enable predictions beyond observed conditions, will be important for the future development of AI for Science aimed at scientific discovery.

Critical weaknesses

I understand that one of the main contributions of this paper is its ability to extrapolate to unseen physical parameter values by prescribing how these parameters enter the governing equations. The authors carefully discuss the non-identifiability and gauge freedoms of physical models in Appendices F.2 and F.5, demonstrating their awareness of these issues. However, it remains unclear to what extent such non-identifiability affects extrapolation predictions and under what conditions the uniqueness of these predictions can be guaranteed. The proposed framework incorporates additional physical constraints, such as energy conservation and the coupling structure between particles and fields, which may resolve these ambiguities. However, it is not clear which constraints contribute to the uniqueness of extrapolation predictions. Furthermore, Proposition 5 assumes that the learned functions g and h coincide with their true counterparts, but whether this assumption can be guaranteed from the training data and the imposed structural constraints is a separate question. In this context, a recent work(https://openreview.net/pdf?id=BDoxdoPBmO), present a framework for evaluating structural uncertainty in physical models based on the rank of the constraint operator. Could the authors clarify whether degrees of freedom that are unidentifiable from the training observations can also affect predictions at unseen parameter values? Possible approaches include analyzing the rank and null space of the observation sensitivity operator or comparing extrapolation predictions from multiple models that fit the training data equally well. Although the extrapolation experiments presented in this paper are interesting, I believe it is important to distinguish whether their success arises from the uniqueness of the extrapolated dynamics or from the selection of an appropriate solution under specific training conditions. This distinction is essential for assessing the reliability and generalizability of the proposed approach.

Questions for authors

Q1. Are the extrapolated dynamics uniquely identifiable from the training data?

Proposition 5 establishes the correctness of extrapolation under the assumption that the learned functions gg and hh coincide with their true counterparts. However, if the training data are restricted to a single parameter value λ0\lambda_0, the transformation (g+δ, h−λ0δ)(g+\delta,\ h-\lambda_0\delta) reproduces the same training dynamics but yields different predictions at unseen parameter values, provided that such a transformation is admissible under the model constraints. Could the authors clarify under what conditions the imposed physical structure guarantees the uniqueness of extrapolated dynamics, even when the individual functions gg and hh are not uniquely identifiable? In particular, which constraints resolve this ambiguity in the charge-reversal and gravity-intervention experiments? Methods for assessing identifiability before training, such as the rank-based framework proposed by a recent work (https://openreview.net/pdf?id=BDoxdoPBmO), may provide a useful reference.

Q2. Is the use of BIC as an approximation to the marginal likelihood justified in this setting?

The conventional BIC approximation relies on regularity conditions, including local parameter identifiability and a nonsingular Fisher information matrix. If non-identifiability remains under the proposed structural constraints, the resulting model may be singular, in which case the conventional BIC is not generally a valid asymptotic approximation to the marginal likelihood (Watanabe, 2013). Moreover, the conservative case with a zero dissipation rate lies on the boundary of the parameter space, potentially introducing additional regularity concerns. Could the authors clarify whether the candidate models satisfy the regularity assumptions underlying BIC? The paper explains that the neural network components are shared across candidate models and that the BIC penalty focuses on differences in structural hypotheses. How is this treatment of model complexity justified from the perspective of marginal likelihood approximation? Given the incorrect model selection observed in the conservative cavity experiment, have alternatives such as WBIC been considered?

Q3. What is fundamentally new compared with existing structure-preserving methods?

The main components of the proposed framework are based on existing methods, while Proposition 5 and the conformal prediction guarantee follow directly from established results. Compared with existing structure-preserving neural networks augmented with the same parameter dependencies and dissipation constraints, what new capabilities or methodological advances does the proposed approach provide?

Q4. How does the proposed method compare quantitatively with the closest existing structure-preserving methods?

The ablation studies and comparisons with unstructured models are carefully conducted. However, comparisons with closely related approaches, such as Dissipative HNNs, port-Hamiltonian neural networks, and GENERIC-based networks, remain largely conceptual. Could the authors provide quantitative comparisons under matched experimental conditions, particularly regarding long-term stability and extrapolation performance?

Reviewer_49ZMOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses an important, and timely, problem of how much inductive bias is needed to be hard-wired into a world model before it can recover known physics. Via numerical experiments, this question is answered for four differing synthetic physical setups. The target behavior is evaluated against a control that removes a single structural component, and thus identifies which component is associated with the resulting capability. Not surprisingly, but perhaps not widely known or appreciated as it should be, it is shown that fully unconstrained models fail to capture dynamics consistent with the correct physics. Conditions of what bias is needed in the four differing cases are presented.

Critical strengths
  • address a fundamental problem of wether unconstrained world models can capture true physics
  • well-written and motivated paper
Critical weaknesses
  • why the choice of those four behaviors? It seems that there could be many possible choices .
  • no advances in theory here
Questions for authors

The results on the four behaviors are observational and their choice seems ad-hoc. It is not clear if the comprehensive results from each of these behaviors can be generalized into a theory across all world model physical predictions. But that may be asking too much at this stage.

Reviewer_9f4HOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

The paper is about how physical structure simulators help the world model recover constitutive relations, transfer to changed physical parameters, and remain bounded over long rollouts.

Its models combine a fixed antisymmetric operator, learned energy functions, and a nonnegative coordinate-damping channel. Experiments cover electromagnetic and shallow-water systems, supplemented by additional physical examples. The main contribution is evaluating these ingredients and their limitations. The paper offers a useful empirical investigation, but the evidence does not sufficiently establish its advance over existing methods, and several headline claims exceed what its guarantees and experiments support.

Critical strengths
  1. A useful scientific question. Separating supplied structure from learned content is valuable. The correctness and consistency inside the world model is an important topic.

  2. Substantial diagnostic effort. Simulator calibration, oracle controls, fresh-data replications, and disclosure of unsuccessful experiments strengthen the study. Distinguishing boundedness, energy fidelity, and trajectory accuracy is appropriate.

  3. Evidence that learned nonlinear content improves prediction. Table 4 reports meaningful rollout improvements over a converged linear constitutive comparator. This is stronger evidence of learning useful content than simply recovering an approximately unit slope.

Critical weaknesses
  1. The empirical contribution is not adequately distinguished from established structured dynamics methods

The principal comparisons are against generic networks and custom ablations. These establish that the selected structural priors help relative to those implementations, but leave unclear what the proposed framework adds beyond a well-implemented existing port-Hamiltonian model. Prior work already includes substantial parts of the claimed capability set. Dissipative SymODEN learns energy, dissipation, and control structure [1]; D-HNN evaluates unseen friction coefficients [2]; and compositional port-Hamiltonian networks preserve energy-exchange structure when assembling learned systems [3]. Appendix N acknowledges these precedents, but a descriptive comparison does not establish the practical distinction.

Moreover, the electromagnetic “non-factored” control changes both parameter dependence and energy-consistent coupling. Its instability therefore does not isolate parameter factorization alone. The fluid experiment helps distinguish these mechanisms, but changes the physical system simultaneously.

A particularly informative comparison would preserve the same energy-derived coupling and numerical solver while replacing only the supplied parameter map with a learned nonseparable map. A matched established PH baseline would also clarify whether the contribution is a new method or a useful evaluation framework.

2.The transfer proposition assumes the crucial identification problem has already been solved

Proposition 5 correctly states that if both components of

G(z;θ)=θg(z)+h(z)G(z;\theta)=\theta g(z)+h(z)

are learned correctly, the supplied map transfers them to new parameter values. However, observing dynamics at one parameter value generally identifies only their combination. For unrestricted component functions, at a training value θ0\theta_0, define

\qquad \widetilde h(z)=h(z)-\theta_0a(z).$$ These components produce exactly the same training generator, but at $\theta'$ their prediction differs by $$(\theta'-\theta_0)a(z).$$ This does **not** invalidate Proposition 5: its assumption explicitly excludes the ambiguity. It does mean that the proposition alone cannot justify identification or counterfactual accuracy from single-setting data. The paper should explain which additional restrictions in each implementation separate the components—for example, known terms, disjoint coordinate channels, symmetry, or sufficiently varied excitation. Appendix F.5 recognizes related identification limits, but the conclusion’s claim that the supplied structures make the constitutive relations identifiable remains too broad. 3. **The “second law” framing is stronger than the guarantee actually established** For Equation (1), with no external drive, $$\dot{E_\theta} = -\sigma_\theta (P\nabla E_\theta)^\top(Pz),$$ which need not be nonpositive. Proposition 3 instead guarantees contraction of a separately specified quadratic diagnostic $Q(Pz)$. The authors explicitly acknowledge this and even provide a counterexample in Appendix C.2. That transparency is a strength. Nevertheless, the heading “The second law holds by construction” and related framing remain misleading: contraction of one coordinate-block diagnostic does not establish an entropy law or passivity with respect to the complete learned energy. The distinction also matters relative to conventional pH damping, $$-R(z)\nabla E_\theta,$$ whose continuous-time energy contribution is nonpositive when $R\succeq0$. Existing pH neural networks explicitly use that construction [1,3]. Likewise, Proposition 4 establishes boundedness for an exact conservative flow under total-energy coercivity and regularity assumptions. It does not establish boundedness of every implemented split update. The empirical stability results remain valuable, but the paper should consistently distinguish: - Exact contraction of the isolated diagnostic. - Conditional continuous-time boundedness. - Observed full-model energy and temperature behavior. ## References [1] Zhong, Y. D., Dey, B., and Chakraborty, A. [Dissipative SymODEN: Encoding Hamiltonian Dynamics with Dissipation and Control into Deep Learning](https://arxiv.org/abs/2002.08860). 2020. [2] Sosanya, A., and Greydanus, S. [Dissipative Hamiltonian Neural Networks: Learning Dissipative and Conservative Dynamics Separately](https://arxiv.org/abs/2201.10085). 2022. [3] Neary, C., and Topcu, U. [Compositional Learning of Dynamical System Models Using Port-Hamiltonian Neural Networks](https://proceedings.mlr.press/v211/neary23a.html). L4DC, 2023.
Questions for authors
  1. What specific capability or finding distinguishes this work from a matched existing pH model? Can one representative experiment establish that distinction?

  2. What prevents the component ambiguity described above in each single-parameter training experiment? Please identify the actual architectural or data assumptions.

  3. What advantage motivates coordinate damping over energy-gradient damping? Can the main claims be consistently limited to the diagnostic that is guaranteed?

  4. Can the authors provide one consolidated results table tied to identified checkpoints and training protocols, including shared-model energy tails, energy supervision, and final scaling verdicts?

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

The paper tackles the problem of learning physical world models able to predict counterfactual queries. The authors study the problem over three physical systems. The methodology is not presented with sufficient detail or presentation rigor, leaving most details to the Appendix.

Critical strengths

Due to the importance of counterfactual reasoning, a research question on how physical world models can represent counterfactuals is interesting and motivates the paper.

Critical weaknesses

The paper does not have a background section. It is therefore extremely difficult to understand and evaluate the content of the paper.

The prose is extremely generic, e.g., "Before any model is trained, each reference simulator must pass conservation tests over the full evaluation horizon"; Which models? How are they trained? Similar considerations apply to further passages.

All the propositions are relegated to the Appendix and not sufficiently discussed in the main body, which only references them. Similarly, the two guarantees mentioned in Section 3.3 are reported in the Appendix and hard to comprehend without a solid background section. In general, the whole Section 3 introduces symbols (zz, qq) and terminology without sufficient detail and without any formalization.

The term counterfactual has a clear definition in the causality literature. A counterfactual measures a causal effect under a given context, intuitively, "Given that I observed Y=yY=y given X=xX=x, what would have been the value of YY if XX were to be x′≠xx' \neq x?". In general, a counterfactual does not coincide with an intervention, since YY might be affected by other variables. This discussion is entirely missing from the paper, and what the authors show looks more like a plain intervention on the physical parameters θ\theta, i.e., the model is trained at one value of θ\theta and evaluated at an unseen value from the same initial state.

Due to my lack of understanding of the methodological section, I have not reviewed the experimental results.

Questions for authors
  • How is a model defined and trained?
  • How is data generated?
  • How are Equations 1, 2, and 3 implemented in a physical world model?
  • Can the authors substantiate the claim that they are studying counterfactual and not interventional queries?
2 comments

I thank the authors for their detailed rebuttal and for their revision of the manuscript.

Section 2.1 now relates the queries to Pearl's definition of counterfactuals and clarifies that, with deterministic simulators and a fully observed initial state, they coincide with interventions; the title and the terminology have been changed accordingly (W4, Q4). The setup has improved as well (Q1, Q2), as has slightly the presentation (W2, Q3).

My presentation concerns remain. Section 2.1 is a problem setup rather than a background section, and the authors adopt notions such as Poisson brackets, Casimirs, coercive energies, and the confining energy class without ever introducing or referencing them (W1). Since ICLR is a general-purpose ML conference, I would have expected at least references. Proposition 1 is still core to Sections 3.3 and 4.1, but stated only in the Appendix (W3). More generally, the main body is still hard to assess, as most results in Section 4 rely on the Appendix. For instance, none of the intervention results in Section 4.3 is reported in a table or figure in the main body; Table 3, cited there, is in App. C.4.

To my understanding of the experiments from a causal perspective, the tested interventions are always of the form do(λ=λ′)do(\lambda = \lambda'), i.e., on a scalar that is constant in the training data. However, λ\lambda is injected into the model through the supplied map λgθ+hθ\lambda g_\theta + h_\theta. Therefore, the model seems able to generalize to interventions "by design", i.e., transfers are explicitly modeled rather than learned. Learning how to learn to model an intervened value, e.g., from data over several values of λ\lambda, would be the interesting causal question, which is not addressed.

Thank you for reading the revision closely and for stating exactly what remains. We revised the paper again (PDF updated) and ran the experiment you describe.

Learning an intervened value from several values of λ\lambda. We agree with your reading. Each test sets a scalar that is constant in training, and the supplied map λgθ+hθ\lambda g_\theta+h_\theta carries the new value into the dynamics, so the transfer is built in. That is the claim: the modeller supplies how λ\lambda enters, and the data supply the content gθg_\theta, hθh_\theta that the map evaluates. §3.5 now says so, and Proposition 6 states when one training value determines that content.

We tested your question on the cavity charge (new App. P.4; design, decision rule and predictions fixed before training).

  • Four ways the charge enters, all else identical: the supplied map αq\alpha q, a learned affine map aq+baq+b, a learned network αnet(q)\alpha_{\rm net}(q), and the charge as an input feature.
  • Training sets of 500 episodes: {+1}\{+1\}, {0.5,1,1.5}\{0.5,1,1.5\}, {−1,+1}\{-1,+1\}.
  • 5 seeds per arm and set, tested at 8 charges from −2-2 to +2+2.

Error relative to the supplied map trained on the same charges, at test charges between and beyond the training charges:

training chargeslearned network: between / beyondcharge as input: between / beyond
{+1}\{+1\}n/a / 2.9 to 80×n/a / 6.4 to 1,200×
{0.5,1,1.5}\{0.5,1,1.5\}1.1× / 1.8 to 7.5×1.4 to 1.8× / 5 to 140×
{−1,+1}\{-1,+1\}1.4 to 2.9× / 1.7 to 15×23 to 220× / 8 to 110×
  • Supplied map. It answers every charge on every set. The extra charges lower its error by at most 9%.
  • Learned network. Several values let it interpolate. Beyond the training charges, its coupling at ∣q∣=2|q|=2 is 12% to 32% off; trained on ±1\pm1, it gives 1.371.37 instead of 22.
  • Charge as input. It interpolates only where training charges are dense, and fails between −1-1 and +1+1.
  • Learned affine map aq+baq+b. This measures Proposition 6 directly.
    • Trained at q=+1q=+1 alone, every model's k(−1)k(-1) equals its initial b−ab-a to three decimals: the data fix only a+ba+b.
    • Trained on {−1,+1}\{-1,+1\}, every model converges to k(q)=qk(q)=q within 3% and answers every test charge within 2× its own in-distribution error. This takes 120 epochs, because 30 are too few for scalars that start far from the answer.
    • Trained on {0.5,1,1.5}\{0.5,1,1.5\}, three of five models converge the same way, and two are still converging at 120 epochs.

So data at several values determine the map where they cover it, as the positivity condition predicts. Beyond that range they determine it only when the class is restricted enough for them to identify it, as the affine map is by two values. The supplied linear map is the strongest such restriction, which is why one value suffices in the paper. Table 1 (§4.3) reports these results next to the single-value ones.

W1, background. The new §2.3 introduces and cites each notion before §3 uses it:

  • the Poisson bracket and its Jacobi identity (Marsden and Ratiu, 1999; Morrison, 1998);
  • Casimirs, with the fluid's water mass as the example (Hairer et al., 2006, ch. VII);
  • coercive energies, and why they bound trajectories (Khalil, 2002);
  • the confining class: a quadratic with positive learned stiffness plus a bounded learned correction;
  • port-Hamiltonian systems (van der Schaft and Jeltsema, 2014).

§2.1 remains the problem setup.

W3, Proposition 1. §3.3 now states the selector's coverage guarantee and its scope, and §4.1 applies it by name.

Results in the main body. The new Table 1 in §4.3 gives every intervention result in the main text:

  • the sign flip on two systems, with the learned-map control and the matched D-HNN and port-Hamiltonian heads;
  • the magnitude, rate, viscosity and gravity transfers;
  • the several-charge results above;
  • the three refusals with their oracles.

The four recovered slopes of the former Table 1 are now a sentence in §4.1, so the appendix tables keep their numbers. With Figures 2 and 3, each behaviour now has its main numbers in the main text.

If these changes and the new experiment address your remaining concerns, we would be grateful if you would consider raising your score.

Comments on the paper

1 comment
Area_Chair_fHivOpenReview ↗

Please engage in the discussion

Dear Reviewers,

As we enter the final stage of the author-reviewer discussion, I would like to encourage everyone to actively engage with the authors’ responses and any additional evidence they have provided.

Please take this opportunity to reconsider your initial assessments with an open mind, particularly where the authors have offered substantive clarifications, detailed arguments, or additional experimental results. The goal of the discussion is not to defend initial positions, but to reach the fairest and most informed evaluation of the submission.

I encourage you to engage directly with the authors on any remaining concerns, clearly distinguishing between issues that have been satisfactorily addressed and those that remain unresolved. Where appropriate, please update your reviews and scores accordingly, or explain why your original assessment still stands.

Your active and constructive participation in these final exchanges is essential to reaching a fair, evidence-based decision.

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.