acceptodds
Withdrawn from ICLR 2027

Causal State Space Models: Routing Dynamic Mechanisms for Non-Stationary Inverse Problems

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

State Space Models (SSMs) provide efficient representations for continuous spatial-temporal fields, yet standard global SSMs suffer from prior entanglement under non-stationary inverse scattering. A localized out-of-distribution (OOD) anomaly shifts the shared transition matrix during online adaptation, corrupting the reconstruction of stationary backgrounds. To address this, we propose the Causal State Space Model (Causal-SSM). By formulating the inverse mapping as a switched linear system via Dynamic Mechanism Routing (DOR), the model partitions continuous data into a library of independent causal mechanisms. During adaptation to physical shifts, our Targeted Mechanism Revision (TMR) confines gradient updates strictly to active anomalous mechanisms while freezing shared parameters and boundary routing. Evaluated on dynamic scattering benchmarks and Institut Fresnel measurements, Causal-SSM achieves robust zero-shot OOD generalization. It successfully assimilates localized anomalies while preserving background fidelity (BRR ), avoiding the severe degradation typical of global-update baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
State Space Models, Inverse Scattering, Online Adaptation, Switched Linear Systems, Out-of-Distribution Generalization, Neural Operators

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zLgBOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper tackles non-stationary inverse scattering with State Space Models (SSMs), noting that adapting a single global transition matrix to a localized anomaly degrades the reconstruction of the stationary background. The authors introduce Causal-SSM, which distributes tokens across a discrete library of KK independent SSM blocks using hard Gumbel-Softmax routing (DOR), and during online adaptation updates only the active anomaly block while detaching the pre-anomaly hidden state (TMR). The model is trained with a variational objective, an episodic meta-learning loss, and contrastive and partition regularizers, backed by theoretical bounds on perturbation drift, boundary approximation, and gradient isolation. Validation is carried out on a synthetic 2D and 3D Helmholtz benchmark and on a transfer experiment using measured Institut Fresnel data.

Critical strengths

S1: Routing tokens to discrete SSM parameter blocks with hard routing, and detaching the hidden state at the anomaly boundary, is a straightforward way to prevent local adaptation updates from leaking into the background representation. In contrast to a global SSM or a soft MoE, where gradients update shared projections across the entire sequence, hard Gumbel routing combined with stop-gradient confines the inner-loop update to the active mechanism while the parallel scan stays intact. The ablation comparing TMR against global adaptation from the same checkpoint illustrates this effect well.

S2: Evaluating synthetic-to-real transfer on the Institut Fresnel measurements is a nice addition beyond the toy synthetic grid. Training on synthetic Born data, calibrating on one measured object, and adapting on held-out inclusions demonstrates that restricting updates to the routed inclusion tokens avoids the background artifacts that global fine-tuning produces.

Critical weaknesses

W1: The paper does not give SSMs a proper introduction, condensing them into a single line (119) with an unexplained formula (e.g., xkx_k is undefined, and Σθ\Sigma_{\theta} appears only in the Appendix). A suitable description can be found in Appendix A.10, yet more than a page of space is left empty in the main text where it could easily have been included. Beyond that, the paper jumps between different concepts and defers even minor details to the Appendix ("preserved in Appendix A.1/A.3", "retained in Appendix A.9"), and BPTT is not defined anywhere in the main text, only in the Appendix. Key parts of the method are missing even in the Appendix: Table 2 indicates that the partition loss (LpartL_{part}) is the main driver of routing, yet its formula is never given anywhere, and Appendix C.11 (line 1393) strangely states that "the supplied manuscript does not provide their exact formulas." The tokenization of the scattered wavefields to match the 12×12 grid is also not described, and there are several reporting errors: UNO and ANFWI-HNO† show identical numbers across Table 1, and the captions of Figures 3 and 6 report different sample IDs and AMI scores from the labels shown inside the plots.

W2: The adaptation setup depends on oracle information that would not be available in a real inverse problem, and the main theorems hold largely by assumption. In Section 3.3, TMR presupposes knowledge of the anomaly interval and of which mechanism is "anomaly-specific", so that htstart−1h_{t_{start}-1} can be detached and the gradients masked. In inverse scattering, the medium and anomaly locations are precisely what we are trying to reconstruct, and the paper acknowledges using oracle regime labels and mutation masks in both training and OOD support adaptation (lines 354, 469, 1455). High-contrast mutations also make up 75% of the training set, so the anomaly mechanism is learned under supervision during training rather than adapted to a genuinely unseen regime. At the same time, Theorem 2 yields zero background gradient precisely because hard routing, state detachment, and a frozen optimizer are assumed up front, and Theorem 3 assumes H(Z∣Z∗)=0H(Z|Z*)=0 and H(Z∗∣Z)=0H(Z*|Z)=0, which already constitutes a bijection when K=K*.

Questions for authors

Q1: Is the approach essentially a hard-routed mixture-of-experts for SSMs, with the experts frozen during adaptation? Also, how are the scattered-field observations tokenized so that each token aligns with one pixel of the 12×12 grid?

Q2: How does an "anomaly-specific mechanism" come to exist in Section 3.3? Do the anomalies appear during training with oracle labels so that this mechanism can be learned, and if so, why are they then called "OOD"? How would TMR function at test time when the anomaly location ImutI_{mut} is unknown?

Q3: Why use phrasing like "retained in Appendix A.9" (line 323) when more than a page of space remains empty in the main text, and why does line 1393 state "the supplied manuscript does not provide their exact formulas" instead of defining LpartL_{part}?

Q4: Why are the numbers for UNO and ANFWI-HNO† identical in Table 1, and why do the captions of Figures 3 and 6 not match the sample IDs and AMI values printed on the figures?

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

This paper proposes Causal-SSM, a state-space architecture for non-stationary inverse scattering. The method combines Dynamic Mechanism Routing (DOR), which assigns tokens to discrete state-space mechanism blocks, with Targeted Mechanism Revision (TMR), which restricts online adaptation to selected mechanisms while freezing shared parameters and inactive optimizer states. The paper also provides conditional theoretical results concerning transition sensitivity, boundary approximation, and parameter isolation. Experiments on synthetic Helmholtz/Born inverse scattering and measured Institut Fresnel data evaluate reconstruction accuracy, OOD adaptation, and background retention.

Critical strengths

The paper explores a meaningful problem in non-stationary inverse scattering. Localized OOD adaptation can interfere with previously learned background representations, and separating anomaly adaptation from background preservation is a practically relevant objective. The proposed combination of mechanism routing and restricted updates provides a reasonable approach to this problem.

Critical weaknesses
  1. The methodological novelty is relatively limited. The main components, including hard expert routing, independent parameter blocks, and selective parameter updates, are closely related to established mixture-of-experts, switched state-space, and modular adaptation approaches. Although integrating these ideas into inverse scattering is potentially useful, the paper does not sufficiently establish what is fundamentally new beyond this combination. In particular, Theorem 1 characterizes a standard limitation of continuous selectors under an assumed discontinuous target, while Theorem 2 largely formalizes the consequences of explicitly isolating parameter blocks and detaching the computational graph. Theorem 3 additionally assumes the conditional-entropy properties needed to establish regime correspondence, rather than showing that the proposed learning objective achieves them. A matched hard-routed, freeze-only baseline without boundary detachment is necessary to isolate the specific contribution of TMR.
  2. Important methodological details are missing, alongside self-contradictory text. Although Table 4 identifies the partition objective as the primary driver of routing alignment, its formula is omitted; bizarrely, Appendix C.11 (line 1383) states in the third person that "the supplied manuscript does not provide their exact formulas", which also directly conflicts with the AI Use Statement on page 9. The transformation from multi-source/receiver scattered-field measurements to spatial reconstruction tokens also requires a clearer operational description. Furthermore, training relies on oracle regime labels and mutation masks, so the current experiments demonstrate supervised regime routing rather than autonomous discovery of physical mechanisms.
  3. Overstated claims and multiple data/figure inconsistencies. Theorem 2 guarantees conditional parameter and pre-anomaly state isolation, but does not prevent altered recurrent states from affecting downstream background predictions. Descriptions of exact background isolation are stronger than what is established, especially given the measured background MSE ratios of 2.8 and 1.9 in Table 8, and global adaptation achieves lower anomaly error than TMR (0.0825 versus 0.1045) in Table 2. Finally, there are concerning reporting inconsistencies: UNO (2022) and ANFWI-HNO (2025) have identical means and standard deviations across all metrics in Table 1, and the captions for Figures 3 and 6 contradict the sample IDs and metric values rendered inside the plots themselves (e.g., sample 76/31 in the Fig. 3 caption vs. sample 115/16 in the plot; sample 101 and dataset-mean AMIs in the Fig. 6 caption vs. sample 37 in the plot).
Questions for authors
  1. What is the specific methodological contribution of TMR beyond conventional selective expert updates or parameter freezing? Could the authors provide a matched hard-routed SSM baseline that updates the same active parameter blocks with identical optimizer-state restrictions but without boundary detachment?
  2. Could the authors provide the exact mathematical definitions of the partition and balance objectives, the contrastive sampling procedure, and the scattered-field-to-token transformation? Additionally, how do the authors explain the phrase "the supplied manuscript does not provide their exact formulas" in Appendix C.11 (line 1383) alongside the AI Use Statement?
  3. How do the authors explain the identical numerical rows for UNO (2022) and ANFWI-HNO (2025) in Table 1, as well as the mismatches between the figure labels and captions in Figures 3 and 6?
Reviewer_g2ZVOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a type of state space model CausalSSM to learn the solution mapping for inverse problems. CausalSSM hard-routes each token to one of the four sets of SSM parameters. On toy inverse scattering problems, the proposed CausalSSM achieves strong zero-shot generalization and background retention.

Critical strengths

None

Critical weaknesses
  1. The empirical evidence is weak. The experiments only consider linearized inverse scattering problems with simple structures (only 12x12 dimensions). The improvement over FNO on the standard NMSE metric is minor.
  2. The theoretical contribution is limited too. The results primarily formalize consequences of the architectural restrictions using standard perturbation, continuity, and differentiation arguments.
  3. It's unclear whether the variational formulation is beneficial. The evaluation only measures the point reconstruction accuracy which cannot justify the use of probabilistic formulation. The paper admits that it does not have calibrated uncertainty. Moreover, the reported ablations do not directly isolate the contribution of the variational formulation.
Questions for authors
  • Can the authors evaluate the proposed method on a nonlinear wave-inversion problem? The problem set up considered in this paper is almost linear. It would be important to show the proposed method works for nonlinear problems. For example, the nonlinear image restoration tasks like phase retrieval in [ 1 ] or the scientific inverse problems in InverseBench [2].
  • Can the authors provide comparison against the classic FISTA-TV inversion algorithm? It's a good conventional baseline to understand the difficulty of the experimental setting.

[1] : Zhang, Bingliang, et al. "Improving diffusion inverse problem solving with decoupled noise annealing." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

[2] : Zheng, Hongkai, et al. "Inversebench: Benchmarking plug-and-play diffusion priors for inverse problems in physical sciences." International Conference on Learning Representations. Vol. 2025. 2025.

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

Related papers

Discussion (0)

Sign in to comment.