In the abstract, you claimed that other methods estimate local graphs independently which makes them vulnerable and not robust, but OGC-MoE does the same during offline training, where you first have labeled data (I’ll detail this point more below), and you train different fθk, so your argument of not using historical knowledge does not hold. You also don’t use historical knowledge to learn your first domain-specific graphs.
One of your arguments (in the rebuttal) might be that you have a testing set where you use a historical set, but many state-of-the-art models that your paper does not cite or discuss learn domain indices from data, so they do not need train and test validation because they learn them.
The paper lacks very important papers in the field: R-PCMCI [1], Castor [2], SpaceTime [3], LIN [4], Fantom [5], On the Identifiability of Switching Dynamical Systems [6], FlowMSM [7], Causal discovery with endogenous context variables [8] and MCD [9]... Please note that state-of-the-art citations and related work are very important scientifically; they help researchers get a clean overview of the different papers. The authors did not compare to these papers in their settings with instantaneous links, non-Gaussian noise, history-dependent noise and heteroscedastic noise. Also, they did not discuss these papers in their manuscript, explaining the differences, where these models outperform their paper, where their paper performs better, or what it offers that is different from these papers. All these papers treat domain indices as latent variables and learn both the regime indices and the graphs. A couple of them assume really challenging settings, including instantaneous links, which are dropped in this paper.
The paper assumes Granger causality, which only takes care of time-lagged relationships. It also returns only a summary graph for every regime, hence we cannot learn which lags change from regime to regime or domain to domain. Also, assuming access to domain indices is a very big assumption.
The idea is poorly motivated. In causality, the objective is to show that this variable X causes that variable Y while estimating the mechanism f of the structural equation. Assuming that the relationship X->Y can be explained by a weighted sum of experts is not well grounded, to me. What is the role of the different experts? The authors used the top 50 or 200 experts in their experiments. Learning a good estimation with different experts, where the information is spread across them, does not help practitioners. Also, I am not aware of any real-world application where you can split the causal relationship into a weighted sum across experts. What would each expert be? Why would I want this? It complicates the interpretability and controllability that causality aims for. Using top-1 routing could be motivated, but top-K with K>1 is, to me, a trick to get great estimation results while trading away the essential spirit of causality: controllability, interventions...
In line 57, you mentioned that state-of-the-art methods “provide limiting support for distinguishing a recombination of known mechanisms from genuinely unseen causal state.” I want to discuss this point: why is this necessary? I cannot find any application for this, i.e. a new domain where, for a variable xi, the structural equation model is a weighted sum of the structural equation models of other regimes. Could you provide real-world examples, with datasets or research papers, confirming that?
I have concerns about the idea of top-K experts for causal discovery. I find it a way of improving estimation, but not very adequate for causal discovery. Happy to discuss!
In line 103, you tried to distinguish your model from causalMoE, but you only describe the difference; you do not motivate why you made this design choice or in which sense it helps your method.
In Methodology, notation and mathematical symbols are not clearly introduced and are used in a confusing way: the authors use non-bold uppercase letters for scalars, vectors, and matrices alike. Using a regular lowercase font for scalars (x), bold lowercase for vectors (x), and bold uppercase for matrices (X) would make the paper easier to read. I understood your method, but I found your methodology section very disorganized and not smooth to read. You do not define Ak as a summary graph for every regime; the reader needs to be able to understand it on its own. You do not clearly state that N is the number of features, nodes..., and you do not explain that you assume access to domain labels during offline training. You mention αk very early at line 137 and only truly explain it at line 244, almost two pages later. You do not define your space ΔK−1 mathematically. Sometimes you put regime indices as superscripts and sometimes as subscripts, which is not rigorous and is misleading. Please unify the notation: A(k),Z(k),e(k)...
I also have concerns about how you decide whether the state is novel or not. Mathematically, any state could be expressed as a weighted sum of existing causal states if these states are a basis for the different graphs (maybe mutually linearly independent, see MCD, Fantom, and On the Identifiability of Switching Dynamical Systems), and somehow your model chooses to create a new state because your states were not expressive enough. The opposite could also happen: your model could use the different causal states to estimate a new state that could never be expressed accurately with the current states. Ablation studies on that and mathematical faithfulness are really needed to trust and use your method.
You introduce a hyperparameter H without discussing how you use it and why it is fixed for all the domains.
Zk were never introduced and explained properly in the paper. I never understood how you compute them or how you optimize them. fθk was also never properly explained: what NN do you use to estimate f, and how are Z and A used inside these models...
You introduce Ik in line 191 without explaining what a valid prediction is. Since you have access to domain labels, what makes a trajectory valid or not?
As you train different fk during offline training, a comparison with MCD, Rhino, CASTOR, FANTOM, Dynotears, PCMCI+... is required to see how you perform against them in their settings.
I acknowledge that the online adaptation is novel; however, I feel that Section 3.3 is somewhat LLM-generated. The algorithm and the equations there are not motivated, and the choices are weird, for example using clip(., ., 20): the choice of 20 is difficult to see as a human choice and as a justified one. The choices of MA and RStd are commonly agent engineering choices. The statistics used in h are weird and not motivated at all. I am not against using LLMs to brainstorm or to run experiments, but I feel the agent was pushed blindly to find a solution for this optimization, so it figured out a way that turns out to be complicated and not motivated. The ct,k equation is non-intuitive for me. Also, agents have the habit of introducing many controllable parameters, which is the case in your paper, like λR,ρz,ρd...
In line 250, you made the choices of 2 experts why this choice, it is never motivated how the model performs with top1 or top 20...?
Why you need calibrated parameters, you introduce new things without motivation and also without goal oriented is just because somehow it worked better.
[1] Saggioro, Elena, et al. "Reconstructing regime-dependent causal relationships from observational time series." Chaos: An Interdisciplinary Journal of Nonlinear Science 30.11 (2020).
[2] Rahmani, Abdellah, and Pascal Frossard. "Causal temporal regime structure learning." AIStats 2025
[3] Mameche, Sarah, et al. "Spacetime: Causal discovery from non-stationary time series." Proceedings of the AAAI conference on artificial intelligence. Vol. 39. No. 18. 2025.
[4] Liu, Chenxi, and Kun Kuang. "Causal structure learning for latent intervened non-stationary data." International Conference on Machine Learning. PMLR, 2023.
[5] Rahmani, Abdellah, and Pascal Frossard. "Flow based approach for Dynamic Temporal Causal models with non-Gaussian or Heteroscedastic Noises." Advances in Neural Information Processing Systems 38 (2026): 8300-8349.
[6] Balsells-Rodas, Carles, Yixin Wang, and Yingzhen Li. "On the identifiability of switching dynamical systems." arXiv preprint arXiv:2305.15925 (2023).
[7] Hulsman, Roel, Carles Balsells-Rodas, and Sara Magliacane. "Identifiable Markov Switching Models with Instantaneous Effects and Exponential Families." arXiv preprint arXiv:2606.02231 (2026).
[8] Günther, Wiebke, et al. "Causal discovery with endogenous context variables." Advances in Neural Information Processing Systems 37 (2024): 36243-36284.
[9] Varambally, Sumanth, Yi-An Ma, and Rose Yu. "Discovering mixtures of structural causal models from time series data." arXiv preprint arXiv:2310.06312 (2023).
Responses to Weakness 1: Our work initially aimed to analyze dynamic dependencies in financial time series. Financial data are continuously generated and nonstationary, while some relationships may persist or recur. We therefore introduced MoE to preserve structures from different historical states, combine historical information when a single expert cannot adequately explain current observations, and accommodate changes through local adaptation and expert expansion.We have clarified the design motivation and scope of our conclusions in the revision.
Responses to Weakness 2 and Question 2: We added sensitivity analyses of historical partitioning and test window length in Sections 4.2.5 and 4.2.6, respectively.
Responses to Question 1: Our initial focus was on the high‑dimensional and nonstationary nature of financial data, making high dimensionality an important context for this work. Offline, we adopt a graph‑conditioned message‑passing recurrent architecture, which reduces parameter duplication compared with constructing an independent prediction network. Online, historical experts’ predictor parameters and structural matrices remain frozen, and routine optimization primarily updates routing weights. Local structural adaptation or new‑expert training is performed only when needed.
We conducted experiments with 64, 128, and 256 variables. As shown below, the model retains good structural recovery performance. Computational costs are reported in Appendix B.11.
Responses to Question 3: We conducted additional experiments retraining the baselines using the current test‑window data. Each baseline uses only observations within the current sliding window, without earlier historical data or historical domain‑partition information.The results are reported in Appendix B.10.
Responses to Question 4: We compared the full method with a setting that prohibits the creation of new experts. As shown in the table below, disabling expansion reduces structural recovery performance on both datasets. Since new experts participate in inference only from subsequent windows onward, these results support the overall benefits of expert expansion for subsequent online structure estimation, complementing the post‑training recovery results on observed segments in Table 5.
The number of variables is fixed at 128 in the runtime experiment below (Appendix B.12).
Responses to Question 5: Weighted expert combination alone cannot guarantee the removal of erroneous edges shared by multiple experts and may retain historical relationships that are no longer valid. However, offline sparsity regularization discourages unnecessary edges, while expert matching and routing reduce the contributions of incompatible experts. When the mixture cannot adequately explain current observations, local adaptation can reduce the relevant edge weights using current data.