- 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 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., is undefined, and 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 () 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 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 and , 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 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 ?
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?