acceptodds
Under review as a conference paper at ICLR 2027

Observationally Equivalent, Interventionally Different: Identifiability of Multiscale World Models

Abstract

A world model can predict every observed coarse trajectory correctly while giving the wrong answer to an intervention on a hidden fast mechanism. We formalize this gap through observational fibers: an intervention target is identifiable exactly when it is constant over all models with the same observational law. This yields a two-point minimax lower bound and a general gauge certificate for non-identifiability. For a slow diffusion driven by a fast Ornstein–Uhlenbeck mechanism, we construct a continuum of microscopic models with the same complete averaged slow-path law but different responses to replacement of the fast conditional mean. We then characterize the missing cross-scale information. Frozen conditional means or a calibrated stationary variance identify the intervention response; one positive-lag covariance together with the variance separates the fast mixing rate from its diffusion scale. A perturbation bound quantifies the instability of intervention transport when the fast mean response is weak. The results distinguish observational prediction, microscopic parameter recovery, and target-specific intervention identification. They also show why architectural clock factorization alone cannot resolve an informational ambiguity in a coarse world model.

Primary area:
causal reasoning
Keywords:
world models, intervention identifiability, multiscale dynamics, observational equivalence, causal representation learning, slow-fast systems, coarse-graining, stochastic dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZNNBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper investigates whether a world model that accurately reproduces observational trajectories can correctly predict the effects of interventions on hidden fast mechanisms. The authors formalize intervention identifiability using observational fibers, showing that an intervention target is identifiable if and only if it is constant across observationally equivalent models. They also derive a two-point minimax lower bound and a gauge-based non-identifiability criterion.

The framework is described using a slow–fast stochastic system in which the fast process follows Ornstein–Uhlenbeck dynamics. The authors construct microscopic models that induce identical averaged slow dynamics but yield different responses to interventions on the fast conditional mean. They then characterize additional measurements sufficient to identify the intervention response, including stationary fast means, variance, and lagged covariance. A perturbation bound quantifies the stability of intervention estimation. Synthetic experiments illustrate the theoretical results, while an additional comparison of single-rate and multirate GRUs examines observational prediction.

Critical strengths
  1. The paper provides a clear mathematical example in which different microscopic mechanisms produce identical observational dynamics but different intervention responses. In particular, the averaged slow dynamics identify the combinations β0=εbr\beta_0=\varepsilon br and β2=εbC2\beta_2=\varepsilon bC^2, whereas replacing CC with a fixed value cc produces the target qc=εbc2=β2c2/C2q_c=\varepsilon bc^2=\beta_2c^2/C^2, which is not identified from the reduced coefficients alone. The exact finite-scale scaling symmetry strengthens this result beyond the averaging-limit example. This provides a concrete illustration of an important limitation of observational world models.

  2. The paper usefully distinguishes identification of an intervention response from recovery of the entire microscopic model. For example, the fast conditional mean identifies CC, making qcq_c identifiable, while the stationary variance can alternatively identify the target through qc=β0c2/rq_c=\beta_0c^2/r. Recovering the fast relaxation rate requires additional temporal information.

Critical weaknesses
  1. The general results in Section 4 are largely direct consequences of the definition of identifiability. Theorem 1 states that an intervention target is identifiable exactly when it is constant on observational equivalence classes; Proposition 1 follows from the standard two-point argument, while Proposition 2 is an immediate consequence of the same equivalence relation. Therefore, these results primarily reformulate established identifiability principles rather than introducing a new theoretical technique. The more substantive contributions concern the specific OU slow–fast model. However, the non-identifiability result follows from the simple parameter ambiguity β2=εbC2\beta_2=\varepsilon bC^2, and the sufficient-anchor results follow from elementary stationary OU moment formulas. Although these derivations are correct and illustrative, it is unclear how substantially they advance existing theory on structural identifiability. In particular, the paper does not provide a general characterization of intervention-identifiable functionals beyond this restricted model class.

  2. Despite the paper's motivation in learned world models, the central experiments use a synthetic system with explicitly specified dynamics and known intervention mechanisms. The anchor experiments initially assume that the reduced slow-dynamics coefficients are known, while the additional joint-estimation experiment uses controlled initial states and independent short trajectories. These conditions are substantially stronger than those encountered when learning a world model from partially observed trajectories. Moreover, the recurrent architecture comparison evaluates only observational prediction through CRPS. It does not evaluate whether learned world models recover the correct intervention response or whether incorporating the proposed anchors improves intervention prediction. Consequently, the experiments show the mathematical construction but provide limited evidence for the paper's broader implications for learning and evaluating multiscale world models. A direct experiment comparing intervention estimates from observationally equivalent learned models, with and without the proposed anchors, would materially strengthen the contribution.

Questions for authors
  1. The observational-fiber characterization and gauge criterion appear to be direct consequences of standard identifiability definitions. Could the authors clarify which general theoretical conclusions are new relative to structural identifiability? Beyond the OU example, does the framework yield a nontrivial characterization of target-sufficient measurements for a broader class of nonlinear or learned dynamical systems?

  2. The non-identifiability construction relies on replacing CC by the same numerical value cc across models with differently scaled hidden fast variables. How is this intervention operationally defined when the fast-state coordinate is unobserved?

  3. Can the authors demonstrate the proposed identifiability gap using learned world models rather than only analytically specified microscopic systems? In particular, do independently trained models with similar observational prediction accuracy yield different estimates of the same intervention response, and do the proposed cross-scale anchors reduce these differences?

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

The paper addresses the problem of characterizing when a coarse observational world model determines the response to an intervention on a hidden fast mechanism. For slow dynamics diffusion there are a continuum of fast Ornstein-Uhlenbeck mechanism consistent with the coarse-grained description. The opposite can also happen, i.e. a mean-response replacement of the fast mechanism produces different coarse responses. Some numerical experiments are carried out to support the conclusions.

Critical strengths

None

Critical weaknesses

This was a dense and not clearly written paper. The motivation and flow of the narrative is abstruse. Perhaps more importantly, the central conclusions are well known in the statistical physics community, and it is not clear what this paper adds. The dynamical renormalization group provides exact conditions for the separation of the multi scale dynamics, and when fast variables have an impact on the universality class of the slow dynamics.

Questions for authors

It would be great if the authors could detail what this paper adds to the huge body of work on the dynamical renormalization group, which does not appear to be referenced at all.

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

The authors formalize the concept of "observational fibers", meaning sets of models that produce identical observational laws. They use it to show that intervention targets are identifiable only if they are constant across these fibers. Using a slow-fast stochastic system (OU process), they prove that a continuum of microscopically different models can be indistinguishable (related to observability in control theory) while responding differently to interventions on the fast mechanism. The work identifies "target-sufficient anchors" (e.g., conditional means and variances) that allow for the identification of intervention responses without requiring full recovery of the microscopic generator.

Critical strengths
  • The use of observational fibers and gauge certificates provides a clean, general mathematical foundation for discussing identifiability.
  • The application to the OU process transforms the abstract theory into a concrete, verifiable example with explicit formulas.
  • The distinction between full system identification and target-specific identification is a valuable contribution to the field of causal representation learning.
Critical weaknesses
  • The numerical studies rely on synthetic data and controlled settings; testing on real-world multiscale datasets or even actual world models would strengthen the claims.
  • The OU process is a standard benchmark, but the generalizability to highly nonlinear fast mechanisms is discussed theoretically but not demonstrated. This makes the relation to world models even weaker. Even without the link, it is not clear how much the results transfer to nonlinear stochastic models.
  • The relation to world models is only tangential, no world models are actually evaluated as a demonstration that the ideas actually apply in practice. This means the study is actually more about slow-fast stochastic systems (like OU processes) rather than world models; I would argue it would have been more focused and thus stronger without the tangent.
Questions for authors
  • Could you discuss how the proposed "anchors" might be acquired in practical settings where frozen fast-state snapshots are not directly available?
  • In the context of world models, what architecture constraints or loss functions could be used to encourage the learning of target-sufficient anchors?
  • How does the stability of the intervention transport (thm. 4) behave when moving away from the linear OU assumption into strongly nonlinear systems?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.