acceptodds
Under review as a conference paper at ICLR 2027

iASIDE: Identifiable, Adaptable, and Separable Interventional Dynamics

Abstract

Learning latent causal dynamics offers an important approach for predicting a dynamics system's response to hypothetical external interventions, especially in systems with high-dimensional observations. This presents two related but distinct challenges: identifiability of the intervention effect in the presence of potential intervention assignment bias, and identifiability of the latent dynamics process that is not directly observed. While the latter challenge persists regardless of the former, it has received limited attention in existing works of intervention response prediction. We bridge this gap with iASIDE, a framework for attaining *identifiable*, *adaptable* and *separable* interventional dynamics. We formulate a latent process with non-stationary parameters and establish its *theoretical* identifiability by conditioning its non-stationary parameters on the variable history. We then address the *algorithmic* challenge of reaching the optimal identifiable solution — caused by the difficulty of separating native and response dynamics from their composite observations — by a separable formulation for both the latent dynamics and the conditional priors of its non-stationary parameters. We realize amortized variational inference of this latent dynamics with feed-forward meta-learning. Experiments on synthetic and real data showed that iASIDE significantly improved the accuracy of intervention response predictions, especially over increased prediction horizons and data heterogeneity.

Primary area:
learning on time series and dynamical systems
Keywords:
intervention estimation, causal, dynamic forecast

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_78NhOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes iASIDE, a latent neural-ODE framework for predicting a dynamical system’s response to sequences of interventions, for example tumour growth under chemo- and radiotherapy. The core idea is to make the latent dynamics depend on a per-system parameter c that is inferred from the observed history, rather than only encoding the history into an initial latent state as existing treatment-response models do. Conditioning c on the history also gives an auxiliary variable in the nonlinear ICA sense, which the authors use to utilise known nonlinear ICA identifiability results. The main novelty is a separable formulation: the dynamics are decomposed into a native term plus additive per-treatment response terms. Each response parameter c^k is inferred only from history segments where treatment k was active. Inference is via amortised variational inference. Experiments on synthetic PK-PD tumour data, pendulum image sequences and MIMIC-III show improved forecasting accuracy over standard baselines, particularly at longer horizons and higher heterogeneity.

Critical strengths

The paper brings identifiable latent dynamics into treatment-response prediction, a gap that existing confounding-focused methods leave open. The separable formulation is intuitive and simple, and experiments cover a range of modalities.

Critical weaknesses

I’m struggling to connect the identifiability theory to the model as implemented. Theorem 1 requires G to be injective in c. However, if treatment k does not occur within a prediction window, c^k has no effect on the outcome, so injectivity seems to fail for exactly the parameters the separable formulation introduces. Can you clarify the conditions under which Theorem 1 applies to c^k? Theorem 2 requires a conditional density p(z_t | z_{t-1}, H_T) with independent components and an invertible emission. The model, though, is a deterministic ODE given c, and for tumour volume and MIMIC the outcome is scalar while z is 16-dimensional. Does Theorem 2 apply to all experiments, or only to the Pendulum experiment? If these points could be addressed adequately I would raise my score

Questions for authors

The separable conditioning of native and response parameters on their own history segments is an intuitive and appealing way to tackle the credit-assignment problem. However, identifying c under the observed treatment patterns seems distinct from identifying the response functions f^k, which is what counterfactual prediction under new plans requires. Can you comment on this distinction?

1 comment

Please see our overall response on Theorem 1 and Theorem 2 which address the two questions raised by the reviewer regarding inactive interventions and the deterministic nature of the ODE when given c\mathbf{c}.

Does identifying c\mathbf c identify the response functions needed for counterfactual prediction

We agree that they are distinct; Theorem 1 identifies only c\mathbf c. The two play different roles in a counterfactual forecast. In our formulation of the data-generation process, fθkf_{\theta^k} is a global function per intervention, learned jointly with the inference of c\mathbf c from the data of all subjects and shared across subjects and plans. Counterfactual prediction under a new plan therefore does not need a new fθkf_{\theta^k}, but the same one evaluated with different timings or combinations of interventions. Furthermore, for the purpose of counterfactual prediction at the observation level (e.g., what will the trajectory of the tumor volume be), what needs to be recovered is actually the mixing function Ga,zT(cK(a))\mathcal G_{\mathbf a,\mathbf z_T}(\mathbf c_{\mathcal K(\mathbf a)}) (which includes both the latent fθkf_{\theta^k}'s and the emission to the data space), along with ck\mathbf c^k that represent the exogenous variable in the underlying causal process that is specific to a subject. The recovery of Ga,zT(cK(a))\mathcal G_{\mathbf a,\mathbf z_T}(\mathbf c_{\mathcal K(\mathbf a)}) across intervention plans observed in the data, as established in Theorem 1, is thus sufficient for counterfactual prediction, such that the patient's response to an alternative counterfactual intervention plan a′\mathbf a' can be predicted by evaluating G^a′,z^T(c^K(a′))\hat{\mathcal G} _ {\mathbf a',\hat{\mathbf z} _ T}(\hat{\mathbf c} _ {\mathcal K(\mathbf a')}) under the new plan a′\mathbf a' at the subject's recovered c^k\hat{\mathbf c}^k's. If a specific intervention plan has never been observed in the data, we will not be able to support their counterfactual prediction within our current theoretical results.

Reviewer_i1QfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is on predicting future outcomes conditioned on history in the latent dynamics framework. It employs theoretical results from iASIDE to reach an identifiable and history-adaptable formulation of latent intervention dynamics, and utilizes variational inference for algorithmically realization. The experiments compares a wide range of seven baselines including causal models, RNNs and transformers. It also presents sota result on real-world dataset (MIMIC-III).

Critical strengths
  • The paper introduces an identifiable formulation for latent dynamics function conditional on history and latent non-stationary exogenous variables.
  • The experiments compares a wide range of seven baselines including causal models, RNNs and transformers. It also presents sota result on real-world dataset (MIMIC-III).
Critical weaknesses
  • The paper needs a more detailed background introduction. For example the formulation of exogenous variables c\mathbf{c} is not clearly introduced at the very beginning. Also, the three key concepts: identifiable, adaptable, and separable, should also be formally defined in somewhere.
  • It seems not all baselines, MSM, RMSN, CRN, CT, ACTIN, SCIP, TESAR-CDE, are reported. For example, In table 2, only 6 out of 7 are reported (no MSM); in table 3, only 5 out of 7 are reported (no MSM and TESAR-CDE). Would you like to complete the missing results? Also, since the paper states its relation with Meta-iLaD, would you also add Meta-iLaD as a baseline? or at least a detailed discussion between Meta-iLaD and the proposed iASIDE.
  • It is unclear to readers what separable formulation means and how it influeces the prediction performance?
Questions for authors
  • What do you mean by algorithmic challenge? is it about time complexity? I didn't see time complexity.
  • What is the formal definition or formulation on identifiable, adaptable, separable, and exogenous variables c\mathbf{c}?
  • What is the setting in MIMIC-III results? like the sample size for training?
  • In section 3.1, could you start with the results on Meta-iLaD? Why their theretical results cannot be directly applied to latent intervention dynamics setting?
  • What separable formulation means? and how it influeces the prediction performance?
2 comments

We thank the reviewer for the detailed comments.

Can the missing baseline results be completed

We have added the missing results: MSM and TESAR-CDE to MIMIC-III (Table 3), MSM to the confounded and counterfactual PK-PD results (Tables 2 and 8), and MSM and SCIP to both Pendulum tables (Tables 6 and 7). These additions did not change the overall conclusion: iASIDE still has the lowest error in every PK-PD setting, and the lowest RMSE and DST on the Pendulum at every horizon, with and without confounding.

How does iASIDE relate to Meta-iLaD, and can Meta-iLaD be added as a baseline

Meta-iLaD (Ye et al., 2026) learns identifiable latent dynamics by conditioning a latent parameter c\mathbf c on a few context samples from the same environment, for systems with native dynamics only. iASIDE extends this by 1) considering the identifiability of ck\mathbf c^k for native and separate intervention dynamics, and 2) the generation of the τ\tau-window outcome, and thus its latent ck\mathbf c^k’s, is conditioned on the history HT\mathcal H_T rather than examples from other systems.

Our revised theorems (summarized in the Overall response T1-T2) have more clearly delineated the changes needed to carry the identifiability result from Meta-iLaD to iASIDE. More specifically, 1) the extension to interventions means that the mixing function from c\mathbf c to the observations now depends on the intervention plan and only the active interventions kk that are present in the given plan; and 2) by conditioning the τ\tau-window generation on the history, the mixing function from c\mathbf c to the observations now also depends on the frames preceding the window (from which the initial state zT\mathbf z_T is computed). Both changes are now reflected in the revised Theorem 1, and this comparison is now stated in Section 3.3 of the revised paper.

As a baseline, Model 3 in our ablation (Table 1) represents the direct application of Meta-iLaD: a single latent ODE with c\mathbf c, conditioned on context examples extracted from the history without separating native or interventional segments. It improves the RMSE over the models without history conditioning (1.21 against 1.64) but recovers the parameters poorly: the MCC is 0.66 for growth and 0.01 for chemotherapy, against 0.94 and 0.76 for iASIDE. As a result, its prediction RMSE is suboptimal compared to iASIDE.

What are the formal definitions of c\mathbf c, identifiable, adaptable and separable

c∈Rdc\mathbf c\in\mathbb R^{d_c} is a subject-specific parameter of the latent dynamics dzt/dt=fθ(zt,at;c)d\mathbf z_t/dt=f_\theta(\mathbf z_t,\mathbf a_t;\mathbf c) (Section 3.1). It is exogenous in that the states and interventions do not affect it; it describes how the dynamics differ between subjects, is assumed constant within a prediction window, and is divided into c0\mathbf c^0 (native dynamics) and ck\mathbf c^k (response to intervention kk).

Identifiable: under the assumptions of Theorem 1, any two models of this form with the same distribution of prediction windows for every history have latent parameters related by a simple function for each ck\mathbf c^k active under the plan. With our Gaussian prior, ck=AkPkc^k+bk\mathbf c^k=A_kP_k\hat{\mathbf c}^k+\mathbf b_k, a component-wise affine transformation (Theorem 1).

Adaptable: the dynamics is not a fixed function, but depends on the history through c∼pϕ(c∣HT)\mathbf c\sim p_\phi(\mathbf c|\mathcal H_T) that enters every step of the rollout; this is in contrast with the common approach where the history enters through the initial state zT\mathbf z_T and then rolls out with a fixed dynamics function.

Separable: the dynamics has an explicit form that includes a native term plus one response term per intervention, fθ=fθ0(zt;c0)+∑katkfθk(zt,atk;ck)f_\theta=f_{\theta^0}(\mathbf z_t;\mathbf c^0)+\sum_k a^k_t f_{\theta^k}(\mathbf z_t,a^k_t;\mathbf c^k) (Eq. 1), and the prior factorizes over the ck\mathbf c^k, with each ck\mathbf c^k inferred only from the history segments that follow intervention kk (Eq. 4). The former is in contrast with the common approach where there is a single latent dynamics without separate native and response components. The latter is in contrast with having a single conditional prior p(c∣HT)p(\mathbf c|\mathcal H_T), as in a direct application of Meta-iLaD, which is Model 3 of our ablation described above.

We have now added an Appendix A to summarize these definitions and concepts in one place, to facilitate the reading of the paper.

AuthorsOpenReview ↗

Response (continued)

What is the algorithmic challenge, and what is the time complexity

By “algorithmic challenge”, we were referring to the optimization difficulty. Theorem 1 alone does not require us to have a “separable” formulation as defined above: it guarantees that c\mathbf c is recovered at the optimum (when observational equivalence is established at the data level), but since the response dynamics are only observed together with the native dynamics, gradient-based training may not reach this optimum. In Table 1, Model 3 is identifiable in theory yet recovers the chemotherapy parameter with an MCC of only 0.01; separating the dynamics and the conditional priors (iASIDE) raises it to 0.76. To avoid confusion, we have now replaced "algorithmic" with "optimization" in the abstract and with "practical" in the introduction.

For computational cost, Table 14 (Appendix I) reports the parameters and the training and inference times on the Pendulum. iASIDE has fewer parameters (19,016) than every baseline except ACTIN, yet trains more slowly (340 s per epoch) because of the iterative ODE solver and the separate meta-model for each ck\mathbf c^k; test-set inference takes 4.2 minutes. The cost per window grows linearly with τ\tau and with the number of history segments aggregated.

What does the separable formulation mean, and how does it affect prediction

As defined above, it has two parts: separate native and response dynamics (Eq. 1), which Appendix B shows is general for binary interventions without multiplicative interactions, and separate priors, with c0\mathbf c^0 inferred from intervention-free segments and ck\mathbf c^k from segments that follow intervention kk (Eq. 4).

Table 1 (PK-PD, heterogeneity 3, τ=5\tau=5) shows its effect. Separating the dynamics alone does not help (Models 2 and 4 against Models 1 and 3), but separating the dynamics and the priors together (Model 5) lowers the RMSE from 1.21 to 1.17 and raises the MCC from 0.66 to 0.94 for growth, 0.84 to 0.90 for radiotherapy and 0.01 to 0.76 for chemotherapy. Appendix H.2 also suggests better sample efficiency with less training data.

What is the MIMIC-III setting

We follow the MIMIC-III setup of Melnychuk et al. (2022): the covariates are 25 vital signs and 3 static features, the outcome is diastolic blood pressure, and the interventions are vasopressors and mechanical ventilation. We have 5000 subjects with a minimum of 30 and a maximum of 60 hours as sequence lengths. These details are now added to Appendix F.3.

Reviewer_2oDpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper aims to estimate future outcomes given hypothetical future actions (interventions), conditional on a latent state, and critically conditional on the current history. The framing enables the model to be parametrized by (i.e. accept as argument) the history of action-outcome pairs, which enable history-adaptive dynamics modelling and estimation of the latent vector in contrast to existing work. The authors demonstrate that under constructed experiments, their approach does resolve issues that existing methods cannot due to this lack of identifiability of the latents and inadaptive parametrizations.

Critical strengths

Addressed gap is real as demonstrated clearly in the constructed/synthetic empirical experiments (Table 1 MCC for chemo is the cleanest). Even if the approach isn't novel from a high level framework and assumption aspects except for this one modelling change, the gap is reasonably motivated as well.

Critical weaknesses

Not having new theoretical results is fine. But: the existing identifiability results for Ye et al. 2026 does NOT apply under the proposed change, meaning the proposed method --- even if adaptable and separable --- is NOT identifiable in theoretical sense. Specifically, eq. (20) and the proof of Theorem 1 is extremely ill-formed.

Less critically, Theorem 2 is ill-posed since according to the paper, dynamics is a deterministic ODE meaning η\eta is degenerate, definitely not twice differentiable. On the empirical side, while the results are convincing under the exact assumption verification setups, existing work (most closely: CT) match or exceed the performance of the proposed method in the real data set experiments.

My main concern is that the identifiability guarantee is effectively asserted without proof, making one of the main arguments irrecoverably wrong.

Questions for authors

None.

1 comment

We thank the reviewer for checking the theory closely.

Theorem 1

Please see our Overall Response T1 for revisions applied to Theorem 1, including the dependence of the mixing function on the initial state zT\mathbf z_T and the corrections to Eq. (20) and the proof.

Does iASIDE improve over existing methods on MIMIC-III

We acknowledge that results from the three top-performing methods, iASIDE, CT, and ACTIN, were not statistically significant. We therefore revised the "state-of-the-art" description in Section 4.3 to explicitly acknowledge this fact, and added a relevant discussion in Section 5 -- Limitations. The newly added MSM and TESAR-CDE did not change the conclusion (RMSE 13.80–38.83 and 10.59–11.13). We did want to bring the reviewer’s attention to the results where we showed that the learned c0\mathbf c^0 captured inter-subject heterogeneity including age (which was not an input) and past vasopressor or ventilation use (Figure 5).

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall response

We thank the reviewers for their careful reading. In response, we stated Theorem 1 more precisely, replaced Theorem 2 with a weaker proposition, and completed the baseline results. Because the changes in Theorem 1 has now helped delineate the theoretical extensions needed to carry that in Ye et al ICML 2026 (Meta-iLaD) to interventional dynamics, we have moved the main theoretical results to Section 3.3 (proof and remarks in Appendix D) and minorly edited the paper throughout to reflect these theoretical contributions. The changes are marked in blue in the revised PDF and the major revisions are summarized below.

T1 (Theorem 1): Impact of inactive interventions and extension of theory from Meta-iLaD (78Nh, 2oDp, i1Qf)

We thank the reviewers for raising these points, which led us to make the statement of Theorem 1 more precise. The mixing function from c\mathbf c to the observations yT+1:T+τ\mathbf{y}_{T+1:T+\tau} in a prediction window depends on 1) the intervention plan, as only the active interventions present in the plan enter the data generation, and 2) the initial state zT\mathbf z_T, computed from the last li+1l_i+1 frames of the history (the initializing frames). We have now revised Theorem 1 to account for both. The resulting guarantee is narrower than the one claimed in the original submission: it holds per intervention, for plans under which the intervention is active, on histories that share the initializing frames.

On "active interventions", the reviewers are right that when an intervention kk is absent throughout a prediction window, the observations in that window carry no information about ck\mathbf c^k. However, this does not undermine our identifiability result because ck\mathbf c^k would not be included in the identification target for that observation window either. Let K(a)\mathcal K(\mathbf a) denote the set of dynamics components active within a prediction window: the native dynamics (k=0k=0 which is always present) together with every intervention kk that is assigned at least once in the window. Since at\mathbf a_t is observed, K(a)\mathcal K(\mathbf a) is known for every window. Because each response term enters the dynamics multiplied by atka^k_t, any term with k∉K(a)k\notin\mathcal K(\mathbf a) vanishes identically throughout the window, so the mixing function G\mathcal G from latent parameters to observations depends on c\mathbf c only through the active sub-vector cK(a)\mathbf c_{\mathcal K(\mathbf a)}. Parameters outside K(a)\mathcal K(\mathbf a) are therefore not part of the generative process for that window; they are unidentifiable from it by construction, but they also have no bearing on its prediction.

The initial state zT\mathbf z_T also determines the mixing function from c\mathbf c to the observations yT+1:T+τ\mathbf y_{T+1:T+\tau}. We assume that zT\mathbf z_{T} is determined from li+1l_i+1 frames of observations preceding the prediction horizon (li≪τl_i \ll \tau; e.g., li=2l_i=2 in our experiments), via a deterministic function Φζ\Phi_\zeta. Restricting Φζ\Phi_\zeta to these frames separates the two roles of the history: the initial state depends only on the frames right before the window, while c\mathbf c depends on the whole history. This allows history where ck\mathbf c^k varies while zT\mathbf z_T stays approximately the same.

Accordingly, we restated T1 for a given plan a\mathbf a, with the noise-free window written as Ga,zT(cK(a))\mathcal G_{\mathbf a,\mathbf z_T}(\mathbf c_{\mathcal K(\mathbf a)}). Assumption 2 now requires the true and learned mixing functions to be injective, with full-column-rank Jacobians, in cK(a)\mathbf c_{\mathcal K(\mathbf a)} only. The identifiability result is now stated for each active intervention and directly for our Gaussian prior, i.e., component-wise affine ck=AkPkc^k+bk\mathbf c^k=A_kP_k\hat{\mathbf c}^k+\mathbf b_k for each k∈K(a)k\in\mathcal K(\mathbf a) (Theorem 1, which now includes the former corollary). With the Gaussian prior, the former assumption 3 on the sufficient statistics always holds and was removed, so the former assumption 4 is now Assumption 3. It asks for 2dk+12d_k+1 histories that end in the same initializing frames and across which the natural parameters of every other active cj\mathbf c^j do not change. As the prior factorizes over the active interventions (Eq. 4), the terms of these cj\mathbf c^j then cancel within each model in the log ratio, leaving only ck\mathbf c^k. With continuous data, this holds approximately for histories close in their initializing frames and in the distribution of the other active cj\mathbf c^j (Remark (v)). The affine map is derived for a given a\mathbf a and zT\mathbf z_T, but depends only on the natural parameters of ck\mathbf c^k at the histories of Assumption 3, so it is shared across plans under sequential overlap (Remark (iii)) and across initializing frames that can follow the same contexts (Remark (iv)).

AuthorsOpenReview ↗

Overall response (continued)

On the proof, 2oDp correctly noted that Eq. (20) used the determinant of the non-square Jacobian of G\mathcal G, and that the step from Eq. (19) was not shown. The revised proof adds this step, with the volume element det⁡(JG⊤JG)−1/2\det(\mathbf J_{\mathcal G}^\top\mathbf J_{\mathcal G})^{-1/2} in place of the determinant, which is defined under the revised assumption 2. The volume terms and the other cj\mathbf c^j then cancel in Eq. (21) under the revised assumption 3, where we also corrected the orientation of the normalizing constants; the remaining steps are unchanged.

These changes have now improved Theorem 1 and allowed us to better delineate the theoretical contribution of iASIDE over the previous results by Meta-iLaD, which the revised paper now states in the Introduction and Section 3.3.


T2 (Theorem 2): Is Theorem 2 well-posed for deterministic latent dynamics, and which experiments does it apply to (2oDp, 78Nh)

We thank the reviewers for these questions. First, we would like to clarify that while the ODE is deterministic when conditioned on c\mathbf c, p(zt∣zt−1,Ht)p(\mathbf z_t|\mathbf z_{t-1},\mathcal H_t) used by Theorem 2 is not because p(zt∣zt−1,Ht)=∫δ(zt−Fθ(zt−1,at−1;c)) p(c∣Ht) dc.p(\mathbf z_t|\mathbf z_{t-1},\mathcal H_t)=\int \delta\big(\mathbf z_t-\mathcal F_\theta(\mathbf z_{t-1},\mathbf a_{t-1};\mathbf c)\big)\thinspace p(\mathbf c|\mathcal H_t)\thinspace d\mathbf c . However, the reviewers are correct that Theorem 2 in the submitted manuscript relied on an unstated assumption: for log⁡p(zkt∣zt−1,Ht)\log p(z_{kt}|\mathbf z_{t-1},\mathcal H_t) to be twice differentiable, the transition needs a density on Rdz\mathbb R^{d_z}, which requires dc≥dzd_c\ge d_z at the least.

Because the identification of c\mathbf c and the forecasting accuracy it enables do not depend on the identifiability of zt\mathbf z_t (Theorem 1 does not use Theorem 2), instead of imposing additional assumptions to substantiate the original Theorem 2, we have now weakened our conclusion on the identifiability of zt\mathbf z_t. The revised Proposition 1 keeps only the invertible emission and the noise condition of Theorem 1, and concludes that z^t=h(zt)\hat{\mathbf z} _ t=h(\mathbf z_t) for a single invertible function h=g^η^−1∘gηh=\hat g_{\hat\eta}^{-1}\circ g_\eta, the same for every time step, history and plan, so the learned dynamics are conjugate to the true ones by hh. The proof follows from the first step of the proof of Theorem 1: trajectories that produce the same noise-free window satisfy gη(zt)=g^η^(z^t)g_\eta(\mathbf z_t)=\hat g_{\hat\eta}(\hat{\mathbf z} _ t) at every step.

Experimentally, Proposition 1 applies to the Pendulum dataset considering zt\mathbf z_t as the position of the pendulum as observed on images; the true latent states zt\mathbf z_t are unknown in PK-PD and MIMIC-III and, since the observation yt\mathbf y_t is one-dimensional (tumor volume in PK-PD and diastolic blood pressure in MIMIC-III), the condition of an invertible mixing function gηg_\eta or g^η^\hat g_{\hat\eta} may be violated. In all datasets, MCC(zt\mathbf z_t) is thus a stronger metric than what the theoretical results establish and is now omitted from the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.