acceptodds
Desk rejected at ICLR 2027

Mamba-B3: Block-Structured, Bounded, and Bijective Frequency SSM for Vision

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision state-space models (SSMs) typically use real-valued diagonal transitions, which efficiently propagate information but cannot represent rotational dynamics that naturally arise in periodic textures, oriented structures, and motion. We introduce Mamba-B3, a vision SSM that combines three complementary modifications to the state transition dynamics. B1 replaces independent real-valued channels with block-structured decay–rotation dynamics and introduces low-rank cross-frequency coupling within each head, allowing frequency components to interact during state evolution. B2 controls the resulting non-normal dynamics through a spectral budget and a block-wise reachability regularizer, providing a sufficient condition for stable state propagation while improving input conditioning. B3 uses bilinear discretization, which, within the one-parameter family of methods, is uniquely able to preserve imaginary-axis modes on the unit circle while maintaining a one-to-one mapping between continuous rotation rates and discrete angles. We evaluate Mamba-B3 on ImageNet-1K, ADE20K, COCO, Something-Something V2, and Breakfast. Across image classification, dense prediction, and video recognition, Mamba-B3 consistently improves over the corresponding vision-Mamba baselines, while systematic component and interaction ablations support the roles of coupling, stability conditioning, and discretization. These results suggest that explicitly modeling and controlling frequency dynamics is a promising direction for more expressive visual SSMs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
state space models, vision backbones, discretization, complex-valued recurrence, stability and conditioning, Mamba, long-range sequence modeling, video understanding

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_hR2oOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper modifies the state transition of vision Mamba models in three ways and applies them to ViM-S for images and VideoMamba-S for video. B1 replaces the real diagonal transition with a block-diagonal decay–rotation generator built from 2×2 blocks with data-dependent decay and rotation rate, plus a rank-RR cross-frequency coupling UV⊤UV^\top inside AA. B2 controls the resulting non-normal dynamics during training through a spectral-norm budget on UV⊤UV^\top, a hinge penalty on the largest eigenvalue of the symmetric part, and a per-block reachability-Gramian conditioning penalty. B3 uses bilinear discretisation, motivated as the unique one-step θ-method that maps the imaginary axis bijectively onto the unit circle, in contrast with ZOH's aliasing of rotation rates. Because of the coupling, the recurrence is evaluated token by token rather than with a parallel scan. With a single model of about 26M parameters, the paper reports 83.0% top-1 on ImageNet at 224² and 84.4% at 384², against 80.3% and 80.8% for its baseline; 49.0 versus 44.8 mIoU on ADE20K; 49.4 versus 46.4 AP on COCO with Mask R-CNN 3×; 70.4% versus 67.6% on SSv2; and 97.1% versus 95.3% on Breakfast. Component ablations are an exclusive 232^3 design on SSv2 and a paired permutation test of the bilinear-by-complex interaction, and the full model runs 29% slower than its baseline.

Critical strengths

The individual theoretical results are correct and combining them for a selective SSM with token-dependent rotation and step size is reasonable. The per-token norm bound does control the time-varying product, unlike frozen-eigenvalue arguments. The SSv2 ablation is also more careful than is typical for backbone papers, with structural null controls in which B2 alone and B2 with B3 reproduce the baseline exactly, and a paired permutation test of the discretisation interaction.

Critical weaknesses

The method is built on a weak baseline and falls clearly behind current vision SSMs

Mamba-B3 is built on ViM-S, an older and weak backbone, and in absolute terms it trails vision SSMs released one to two years ago: VSSD-T reaches 83.7% and M2D-T 84.0% at 4.5 to 4.7 GFLOPs, against 83.0% at 5.8 GFLOPs. The paper sets this gap aside by restricting its claims to a self-defined class of isotropic backbones, and on dense prediction Sec. 5.3 reports the margins over VMamba-T and VSSD-T but does not mention that DAMamba-T, listed in the same Table 3, beats Mamba-B3 on every metric, with 50.3 versus 49.0 mIoU and 50.4 versus 49.4 AP. Only one model of about 26M parameters is reported, with no larger variants and no transfer to a stronger backbone, and throughput and memory are compared only with the paper's own baseline, showing a 29% throughput loss and 21% more memory.

The novelty is narrower than claimed, and the gains are not attributed to the novel part.

The rotation blocks come from Mamba-3, as the paper acknowledges, and bilinear discretisation is the standard choice of S4 and DSS, so the main new element is the low-rank coupling inside AA. However, data-dependent diagonal-plus-low-rank transitions that couple state channels are exactly what DeltaNet [1], Gated DeltaNet [2], RWKV-7 [3] and DeltaProduct [4] use, and none is discussed, although the paper cites Grazzi et al. (2025), which studies this family; these works also train such transitions with chunkwise-parallel algorithms, which calls into question the claim that the coupled recurrence must run token by token. For B2, the symmetric-part constraint follows Lipschitz RNNs [5], and Gramian-based regularisation during SSM training was proposed in [6]. The experiments then do not isolate the coupling: no model uses the rotation blocks with R=0R=0, the coupling norm is bounded by amin⁡=1a_{\min}=1 by construction, the configuration uses a state of 128 per head against 16 for ViM-S and 0.7 more GFLOPs, and the parameter-matched baseline matches neither state size nor FLOPs. The rationale for B3 is also contradicted by the paper's diagnostics, since Table 5 reports mean rotation angles of only 0.19 to 0.29 rad, far from the aliasing regime near π\pi that B3 is meant to fix.

References

[1] S. Yang, B. Wang, Y. Zhang, Y. Shen, Y. Kim. Parallelizing Linear Transformers with the Delta Rule over Sequence Length. NeurIPS, 2024.

[2] S. Yang, J. Kautz, A. Hatamizadeh. Gated Delta Networks: Improving Mamba2 with Delta Rule. ICLR, 2025.

[3] B. Peng, R. Zhang, D. Goldstein, E. Alcaide, X. Du, et al. RWKV-7 "Goose" with Expressive Dynamic State Evolution. arXiv:2503.14456, 2025.

[4] J. Siems, T. Carstensen, A. Zela, F. Hutter, M. Pontil, R. Grazzi. DeltaProduct: Improving State-Tracking in Linear RNNs via Householder Products. NeurIPS, 2025.

[5] N. B. Erichson, O. Azencot, A. Queiruga, L. Hodgkinson, M. W. Mahoney. Lipschitz Recurrent Neural Networks. ICLR, 2021.

[6] P. Schwerdtner, J. Berman, B. Peherstorfer. Hankel Singular Value Regularization for Highly Compressible State Space Models. NeurIPS, 2025.

Questions for authors

(1) Please add the rotation-only transition with R=0R=0 on SSv2 and ImageNet, with the same B2 and B3 settings, and report the distribution of ∣ωhΔt∣|\omega_h\Delta_t| over tokens and layers. what fraction ever exceeds π/2\pi/2?

(2) What are the state dimension and recurrent FLOPs of the baseline and of each B1 variant? Please add a baseline matched in both state size and FLOPs.

(3) How does the coupled transition differ from the diagonal-plus-low-rank transitions of DeltaNet, Gated DeltaNet, RWKV-7 and DeltaProduct, and can it be computed with their chunkwise-parallel algorithms?

(4) Does B1+B2+B3 improve a strong current backbone such as VSSD-T, M2D-T or DAMamba-T under its original recipe, and what are the throughput and memory of Mamba-B3 compared with these models on the same device?

6 comments
AuthorsOpenReview ↗

Clarifications About the Contributions and New Results (1/2)

We thank the reviewer for the detailed review and the references. Below, we address each comment in detail.

"The method is built on a weak baseline and falls clearly behind current vision SSMs"

Scope and choice of backbone

The paper evaluates a state-transition design and does not claim state-of-the-art accuracy. ViM-S was chosen for three reasons:

  1. It is a plain isotropic backbone in which the selective SSM layer is the only sequence mixer, so every difference between rows is attributable to the transition.
  2. The model's one-dimensional bidirectional scan uses the standard Mamba layer, which VideoMamba extends, so one implementation covers image and video tasks.
  3. It keeps the full token sequence at every layer (196 tokens at 2242224^2, 576 at 3842384^2), the long-scan regime the stability analysis targets.

Within that scope the effect is large for a change confined to the transition: +2.7 top-1 on ImageNet-1K, +4.2 mIoU on ADE20K, +3.0 AP on COCO and +2.8 on SSv2 over the same backbone, at 0.4M added parameters and with a parameter-matched baseline 2.4 points below (Tables 2 to 5).

Absolute standing

We agree that Mamba-B3 trails newer backbones. Section 5.2 reports the gaps to VSSD-T (0.7) and M2D-T (1.0). Section 5.3 lists DAMamba-T in Table 3 but omits it from the text; the text will state that DAMamba-T exceeds Mamba-B3 on both dense tasks (50.3 against 49.0 mIoU, 50.4 against 49.4 AP). The revised paper will state the same to compare transparently against the latest published work. The paper reports one model size, about 26M parameters, matching ViM-S and VideoMamba-S. Larger variants were not trained, and the revised paper will state this as a limitation.

Transfer to a stronger backbone (Question 4)

We agree that transfer to a stronger backbone would strengthen the evidence for general applicability. A controlled comparison requires training both the original backbone and its modified counterpart under the same full ImageNet-1K recipe. Given our computational resources, we cannot complete these experiments before the discussion closes. Transfer also requires architectural adaptation rather than a direct layer replacement, particularly for backbones with different recurrence structures. Meanwhile, we will benchmark the public VSSD-T, M2D-T, and DAMamba-T checkpoints alongside Mamba-B3 on the same device for throughput and memory, under the protocol of Table 1 (batch 128, fp16, 2242224^2 input). These measurements require inference only and will be included in the revision uploaded during the discussion period.

"The novelty is narrower than claimed, and the gains are not attributed to the novel part"

Delta-rule transitions (Question 3)

These works are directly relevant and will be discussed in the updated related works section. Mamba-B3 differs in three respects:

  1. Its base is a continuous-time block-diagonal decay-rotation matrix, where DeltaNet, Gated DeltaNet and RWKV-7 have real spectra and DeltaProduct obtains rotations from products of reflections.
  2. Our low-rank term is a learned coupling shared across tokens, where the delta-rule family uses key-dependent updates.
  3. Mamba-B3 is discretized under a token-dependent step.

We agree that token-by-token evaluation is not forced by the coupling: the chunkwise algorithm of [1] has been adapted to token-dependent diagonal-plus-low-rank transitions [2], and we withdraw the argument of Section 5.1 that the coupled path must be sequential. We have not implemented a chunkwise kernel for the block-diagonal base. We will cite [3] for the symmetric-part constraint and [4] for Gramian regularization, and correct the Related Work sentence that describes control-theoretic analyses as diagnostics only.

References

[1] S. Yang, B. Wang, Y. Zhang, Y. Shen, Y. Kim. Parallelizing Linear Transformers with the Delta Rule over Sequence Length. NeurIPS, 2024.

[2] S. Yang, J. Kautz, A. Hatamizadeh. Gated Delta Networks: Improving Mamba2 with Delta Rule. ICLR, 2025.

[3] N. B. Erichson, O. Azencot, A. Queiruga, L. Hodgkinson, M. W. Mahoney. Lipschitz Recurrent Neural Networks. ICLR, 2021.

[4] P. Schwerdtner, J. Berman, B. Peherstorfer. Hankel Singular Value Regularization for Highly Compressible State Space Models. NeurIPS, 2025.

AuthorsOpenReview ↗

Clarifications About the Contributions and New Results (2/2)

Isolating the coupling (Question 1)

New controls on SSv2 (VideoMamba-S, 16×224216 \times 224^2, bilinear, mean of 5 runs per new cell, top-1):

No couplingLow-rank coupling
Real diagonal, B2 off67.968.3
Rotation blocks, B2 off68.669.4
Rotation blocks, B2 on68.770.4

Coupling adds 0.8 without B2 and 1.7 with the B2 and B3 settings the reviewer asks for, against 0.4 on a real diagonal. The coupling contributes most when it is bounded and conditioned. Rotation alone accounts for 0.7 of the 1.5 that B1 gains without B2, and the introduction will be reworded accordingly. We did not repeat this control on ImageNet-1K due to time constraints.

Coupling budget

The bound ∥UV⊤∥2≤amin⁡\lVert UV^\top \rVert_2 \le a_{\min} applies only when B2 is active; the B1-only and B1+B3 rows train without it (ρ=1.03\rho = 1.03 and 0.990.99 in Table 5).

The existing ablation does isolate B2, B3 and their interactions with B1, which the reviewer graciously credits: the null controls reproduce their references exactly, the sub-ablation separates the spectral projection from the conditioning term, and the full model exceeds every pair.

State size and FLOPs (Question 2)

The reviewer is right that the two layouts differ, the value of per-head state of N=128N = 128 only applies to the variants with B1 active, which use 12 heads, each with N=128N = 128 and H=64H = 64 rotation blocks. The baseline and the two variants without B1 use the unmodified ViM-S layer, with a state of 16 per channel and no heads, which is why the baseline matches ViM-S at 26.0M parameters and 5.1 GFLOPs. We will correct the Configuration paragraph to state both layouts. The measurements quoted at N=128N = 128 in Section 3.2 and in the appendix refer to the B1 state and are unaffected. Matching the total parameter count does not close the gap: the parameter-matched baseline reaches 80.6 against 83.0, at 5.5 against 5.8 GFLOPs.

Mean rotation angle and B3

Thank you for raising this concern. We would like to gently clarify that we do not see a contradiction between the paper's diagnostics and the rationale for B3, for the following reasons:

The caption of Table 5 describes ∣θ∣\lvert \theta \rvert as a location statistic for the rate distribution, and a mean of 0.19 to 0.29 does not bound its tail. The paper's evidence for B3 is the interaction of Table 6, where bilinear adds 0.3 on a real-diagonal state and 1.2 on a complex-block state, and that result stands independently of the mean angle.

Over all layers and tokens of the ZOH model, ∣ωΔ∣\lvert \omega \Delta \rvert has a median of 0.22, a 95th percentile of 1.85, a 99th percentile of 3.40, a 99.9th percentile of 5.60, a maximum of 8.4 with 1.5% of observations exceeding π\pi, and 6.5% exceeding π/2\pi/2.

Conclusion

The new results separate rotation from coupling, correct the stated state layout, quantify the efficiency trade-off, and position the method against the delta-rule family; a transfer to a stronger backbone was not run. We appreciate the reviewer's comments and are committed to addressing these points in the revised version. We believe our contributions are valuable to the community researching SSM/Mamba Models and kindly request the reviewer to reconsider their initial score in light of these improvements.

AuthorsOpenReview ↗

Thanks and Follow-Up

We would like to thank you again for your careful review and constructive suggestions. We hope our responses have addressed your concerns regarding the paper. Please let us know if you have any further questions or if any additional information is required.

Revised manuscript to address comments and concerns has been uploaded as promised.

Reviewer_YPD3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Mamba-B3, which modifies the state transition of vision Mamba models in three ways. First, it introduces block-structured rotation dynamics with low-rank coupling between frequency components. Second, it applies spectral constraints and a block-wise reachability regularizer to control the transition dynamics. Third, it replaces ZOH with bilinear discretization. The method is evaluated on image classification, detection, segmentation, and video recognition tasks, with improvements reported over the original Mamba backbones.

Critical strengths

The paper studies the state transition itself rather than introducing another scanning strategy. The authors also provide component ablations and report actual throughput and memory usage, which help clarify the cost of the proposed modifications.

Critical weaknesses

The main components are largely established techniques. Block-structured rotation dynamics have already been explored in Mamba-3, diagonal-plus-low-rank state transitions were studied in S4, and bilinear discretization is a classical approach in state-space modeling. The spectral constraints and reachability regularization are also based on existing control-theoretic tools. The paper combines these ideas, but it is not sufficiently clear what new modeling principle is introduced beyond their integration. In particular, the claimed uniqueness of bilinear discretization holds only within a restricted family of numerical schemes and is a known mathematical property.

The reachability regularizer is computed using isolated rotation blocks with frozen parameters and without cross-frequency coupling. However, the proposed model uses coupled, input-dependent transitions. Better conditioning of individual blocks does not necessarily imply better information propagation through the complete system. The paper relies on this argument to explain the improvements on long-horizon video tasks, but does not directly validate it. Similarly, the mathematical properties of bilinear discretization do not establish that avoiding frequency aliasing is responsible for the reported accuracy improvements.

The low-rank coupling breaks the efficient associative scan and requires sequential recurrence. Table 1 shows that throughput decreases from 1320 to 940 images/s, while memory usage increases from 8.5 GB to 10.3 GB. This is a considerable cost for a method built on efficient state-space modeling. The paper does not provide comparisons under matched inference time, making it difficult to determine whether the improvement comes from a better architecture or simply from more expensive computation.

Questions for authors
  1. What is the main technical contribution beyond combining the block-structured transitions of Mamba-3, the low-rank transition structure of S4, and classical bilinear discretization?

  2. Why should the reachability of isolated rotation blocks be considered evidence of improved information propagation in the full coupled, input-dependent model? Can the authors provide direct measurements on the complete system?

  3. How often do the learned frequencies actually enter the aliasing regime under ZOH? Without this analysis, it is difficult to attribute the benefit of bilinear discretization to the proposed frequency argument.

  4. Does Mamba-B3 still outperform the baselines when inference latency or throughput is matched?

3 comments
AuthorsOpenReview ↗

Clarifications About the Contributions and New Results (1/2)

We thank the reviewer for the detailed feedback. Below, we address each comment.

"The main components are largely established techniques" (Weakness 1, Question 1)

We agree that the individual tools are classical, and the paper states this: Section 3.3 identifies Eq. (7) as the classical unit-modulus property of the θ\theta-method, and Appendix A.1 presents each result as a restatement of a known one.

We do not agree that no new modeling principle is introduced. The contribution is a specific transition parameterization and its combination with a sufficient stability constraint. Coupling frequency across blocks raises the capacity of rotational dynamics. Under the block-rotation structure the coupling is the only term that can expand the state, so a budget on it alone bounds the operator. With rate and step size both token-dependent, bilinear discretization keeps the rate identifiable.

S4 introduced its low-rank term to make the HiPPO matrix diagonalizable under a time-invariant step, and diagonal-only variants later matched it. In Mamba-B3 the low-rank term couples rotation blocks whose rates and step size are token-dependent.

The submitted paper already tests the components as an interacting system. In the exclusive ablation of Table 5 (five seeds per row) the full model exceeds every pair (70.4 against 69.4), B1 alone is the only unstable configuration (ρ=1.03\rho = 1.03) and returns below one whenever B2 is added, and the two null controls reproduce their references exactly. Table 6 shows that bilinear is worth 0.3 on a real-diagonal state and 1.2 on a complex-block state. These dependencies are the evidence that the three components operate as one design.

A new control tests the one dependency that design did not separate.

New control separating rotation from coupling

SSv2, VideoMamba-S at 16×224216 \times 224^2, bilinear, mean of 5 runs per new cell on H100, top-1:

No couplingLow-rank coupling
Real diagonal, B2 off67.968.3
Rotation blocks, B2 off68.669.4
Rotation blocks, B2 on68.770.4

Without B2, coupling adds 0.8 on rotation blocks against 0.4 on a real diagonal. With B2 it adds 1.7. The coupling is worth about twice as much when the state rotates, and it contributes most when it is bounded and conditioned. Rotation alone accounts for 0.7 of the 1.5 that B1 gains without B2, and we will reword the stated distinction from Mamba-3 accordingly.

The stability result specializes classical facts (Appendix A.1) to a non-normal, token-dependent operator. Eigenvalue constraints such as those of LRU and StableSSM bound a normal operator and do not bound a non-normal one (Section 2). Because the symmetric part of each rotation block is ahIa_h I, a budget on ∥UV⊤∥2\lVert UV^\top \rVert_2 alone gives ∥Aˉt∥2≤1\lVert \bar{A}_t \rVert_2 \le 1 at every token and step size under ah≤−amin⁡a_h \le -a_{\min} (Lemma 2, Proposition 2), with nothing checked per token.

"Better conditioning of individual blocks does not necessarily imply better information propagation through the complete system" (Weakness 2, Question 2)

We agree. Section 3.2 states that the penalty "conditions that block and not the coupled scan". We therefore measured the complete system: a finite-horizon reachability Gramian of the deployed recurrence, with the learned coupling and token-dependent transitions, along 5 validation sequences.

Conditioning raises the effective rank of the coupled Gramian by 25%. The ratio λmin⁡/trace⁡\lambda_{\min} / \operatorname{trace} stays small in every model and improves by less than 3×3\times. On Breakfast, the final state is 1.24×1.24\times more sensitive to final-quartile tokens in the full model than in the B2-inactive variant, and about 1.0×1.0\times in the earlier quartiles. This is a frozen-trajectory analysis, so we present the Breakfast result as consistent with the conditioning hypothesis.

The conditioning argument also produced a prediction that held: the two Breakfast variants stay within 0.4 points through 75% of frames and separate only in the final quartile (Table 4).

AuthorsOpenReview ↗

Clarifications About the Contributions and New Results (2/2)

"The mathematical properties of bilinear discretization do not establish that avoiding frequency aliasing is responsible for the reported accuracy improvements" (Weakness 2, Question 3)

We agree that the mathematical properties of bilinear discretization alone do not establish the cause of the accuracy improvements. We additionally quantified the frequency distribution of the trained ZOH model by measuring ∣ωΔ∣\lvert \omega \Delta \rvert across all layers and tokens on SSv2 validation. The distribution has a median of 0.22, a 95th percentile of 1.85, a 99th percentile of 3.40, a 99.9th percentile of 5.60, and a maximum of 8.4 with 1.5% of observations exceeding π\pi.

MetricValue
Median0.22
95th percentile1.85
99th percentile3.40
99.9th percentile5.60
Maximum8.40
Fraction exceeding π/2\pi/26.5%
Fraction exceeding π\pi1.5%

These measurements establish that a portion of the learned base-block rotation increments extends to a measurable high-frequency tail, providing empirical support for the aliasing motivation behind B3. This diagnostic complements Section 5.6, which independently demonstrates that the benefit of bilinear discretization depends on the state structure.

"The low-rank coupling breaks the efficient associative scan and requires sequential recurrence" (Weakness 3, Question 4)

The cost is as the reviewer states: 940 against 1320 img/s and 10.3 against 8.5 GB. ImageNet-1K top-1 against measured throughput on one H100 (batch 128, fp16):

ModelTop-1img/s
Baseline, real diagonal80.31320
Mamba-B3, R=1R = 182.41130
Mamba-B3, R=2R = 283.0940
ViM-B (98M, published top-1)81.9351

Rank 1 retains 78% of the ImageNet gain at 86% of baseline throughput, and Mamba-B3 exceeds ViM-B by 1.1 points at 2.7×2.7\times its throughput. We have not trained a latency-matched baseline. The token-by-token evaluation is a property of our implementation and is not forced by the coupling: chunkwise algorithms have been implemented for token-dependent diagonal-plus-low-rank transitions [1]. We have not yet written such a kernel for the block-diagonal base.

The paper reports this cost in measured throughput and memory because FLOPs do not capture it (Table 1), and it shows that rank controls the cost: R=1R = 1 recovers half of the lost throughput. Matching the total parameter count does not close the gap: the parameter-matched baseline reaches 80.6 against 83.0 (Table 2).

Conclusion

The new controls separate rotation from coupling, measure conditioning on the complete recurrence, quantify how often rates fold, and place the model on an accuracy versus throughput curve. We believe they address the concerns on soundness and contribution, and we kindly request the reviewer to reconsider the initial score.

References

[1] S. Yang, J. Kautz, A. Hatamizadeh. Gated Delta Networks: Improving Mamba2 with Delta Rule. ICLR, 2025.

AuthorsOpenReview ↗

Thanks and Follow-Up

We would like to thank you again for your careful review and constructive suggestions. We hope our responses have addressed your concerns regarding the paper. Please let us know if you have any further questions or if any additional information is required.

Revised manuscript to address comments and concerns has been uploaded as promised.

Reviewer_WGbeOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper replaces the selective state-space layer of ViM-S and VideoMamba-S with a transition combining damped two-dimensional rotations and a low-rank cross-block correction, a coupling norm budget together with stability and block-reachability losses, and bilinear discretization. Its central argument is that these choices work together: coupling increases expressivity but introduces non-normal dynamics, the budget controls their growth, and the bilinear map preserves distinct continuous oscillation rates within a specified discretization family.

Critical strengths
  1. The paper gives a coherent dynamical interpretation of its architectural changes. In particular, controlling the symmetric part of a non-normal generator is better justified than relying on eigenvalue magnitudes alone.

  2. There is meaningful experimental breadth. Improvements occur in classification, dense prediction, and video recognition, and Table 2 includes a parameter-matched baseline.

  3. The writting is easy to follow, and figures are neat.

Critical weaknesses
  • The B1 mentioned in this paper simultaneously introduces rotations and the low-rank correction. Thus B1-only versus baseline, and the "Complex block" rows of Table 6, cannot distinguish the benefit of oscillatory channels from the benefit of coupling them. The parameter-matched ImageNet baseline controls capacity, but not this distinction. I would like to see a trained rotation-only model with U = V = 0, followed by the same model with coupling, a real-diagonal-plus-low-rank variant would further clarify the interaction. This is especially important because the paper identifies coupling inside the transition, rather than complex dynamics themselves, as its primary difference from Mamba-3.

  • This paper acknowledge that the regularizer removes coupling and freezes one token. However, subsequent explanations of late evidence rely heavily on this surrogate. Even with frozen block-diagonal dynamics, full reachability does not follow from each block being reachable: two identical rotation blocks driven by the same scalar input can each have rank-two local reachability while the combined four-dimensional system has only a two-dimensional reachable subspace. Time-varying input maps and coupling create additional discrepancies.

  • Equation (5) uses an unnormalized minimum singular value. Scaling every input map by a positive factor scales this singular value by that factor and reduces the penalty. In the SSM as written, inversely scaling the output map can preserve predictions. Consequently, a larger reported singular value need not mean that previously weak directions became more accessible relative to other directions. This is a possible degeneracy of the objective, not evidence that training actually exploits it. Please report input/output projection norms and a scale-normalized reachability measure, or compare against a normalized objective.

Questions for authors
  1. What accuracy and measured throughput do rotation-only transitions achieve before low-rank coupling is added? Does coupling still help relative to this directly relevant baseline?
  2. How are the real-diagonal rows grouped to produce the identically rank-one diagnostics in Table 5? Are paired decay entries tied, or is the conditioning calculation disabled?
  3. Does the conditioning penalty improve a normalized Gramian or the reachability of the actual coupled, time-varying scan after controlling for input-map magnitude?
2 comments
AuthorsOpenReview ↗

Clarifications About the Contributions and New Results

We thank the reviewer for the careful reading and the precise counterexample. Below, we address each comment.

"The B1 mentioned in this paper simultaneously introduces rotations and the low-rank correction" (Weakness 1, Question 1)

We agree, and we ran the controls the reviewer describes. All four transitions on SSv2 under one protocol (VideoMamba-S, 16×224216 \times 224^2, bilinear, B2 off, mean of 5 runs per new row on H100).

TransitionTop-1Throughput (img/s)
Real diagonal67.91294
Real diagonal with low-rank coupling68.31026
Rotation blocks only (U=V=0U = V = 0)68.61203
Rotation blocks with coupling (B1)69.4942

Coupling adds 0.8 over the rotation-only model, the directly relevant baseline, and 0.4 over the real diagonal; rotation alone adds 0.7. Coupling still helps relative to rotation only, and the rotation-only model keeps most of the baseline throughput, so the cost of B1 lies in the coupled recurrence. With B2 and B3 active, the rotation-only model reaches 68.7 against 70.4 for the full model.

"Subsequent explanations of late evidence rely heavily on this surrogate" (Weakness 2)

We agree, and the counterexample is correct. With a shared scalar input, the frozen uncoupled system is reachable only if every block is reachable and the block eigenvalues are pairwise distinct; the per-block penalty measures the first condition only. We therefore measured the deployed recurrence, with the learned coupling and token-dependent transitions, using a finite-horizon Gramian along 5 validation sequences.

Conditioning raises the effective rank of the coupled Gramian by 25%, and λmin⁡/trace⁡\lambda_{\min} / \operatorname{trace} stays small in every model, improving by less than 3×3\times. The improvement on the coupled scan is modest. This is a frozen-trajectory measurement. We describe the penalty as a per-block regularizer, and Section 5.4 will say that the Breakfast result is consistent with the conditioning hypothesis.

"Equation (5) uses an unnormalized minimum singular value" (Weakness 3, Question 3)

The normalized measurement shows that the conditioning term improves block reachability beyond input-map scale: σmin⁡/∥Bˉh∥2\sigma_{\min} / \lVert \bar{B}_h \rVert_2 rises by 1.9×1.9\times once the 1.25×1.25\times scale factor is removed. The sub-ablation of Table 5 attributes this change to the conditioning term. The spectral projection alone leaves σmin⁡\sigma_{\min} at 0.031 against 0.029, and the conditioning term alone raises it to 0.069 and Top-1 from 69.4 to 69.9. The penalty weight is small, λcond=0.01\lambda_{\text{cond}} = 0.01 (Configuration), which limits the pressure toward scale growth. The per-block form is what makes reachability usable as a training objective here, since the full-state objective collapses at realistic state sizes (Section 3.2).

"How are the real-diagonal rows grouped to produce the identically rank-one diagnostics in Table 5?" (Question 2)

The paired decay entries are not tied, and Lcond\mathcal{L}_{\text{cond}} is not computed for real-diagonal variants, which contain no rotation blocks (Configuration, Section 4). The rank-one entries in Table 5 are the analytic value of Proposition 3(iv) for a block without rotation, which is a scaled identity with collinear reachability columns and σmin⁡=0\sigma_{\min} = 0. Two distinct decays give rank two, since

det⁡[B,AB]=b1b2(d2−d1).\det [B, AB] = b_1 b_2 (d_2 - d_1).

The entries are not per-run measurements and will be marked N/A for clarification.

Conclusion

The controls isolate the coupling, measure reachability on the coupled scan, and separate conditioning from input-map scale. We believe they resolve the remaining concerns and kindly request the reviewer to consider them in the final assessment.

AuthorsOpenReview ↗

Thanks and Follow-Up

We would like to thank you again for your careful review and constructive suggestions. We hope our responses have addressed your concerns regarding the paper. Please let us know if you have any further questions or if any additional information is required.

Revised manuscript to address comments and concerns has been uploaded as promised.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers and Revision Schedule

We thank the Reviewers for their detailed and constructive reviews. Individual responses are posted under each review.

The revised paper has been uploaded, with all changes marked in blue. It includes the new results, corrections and clarifications referenced in our individual responses, and the main text remains within the 10-page limit.

We respectfully ask all reviewers to consider these responses and the revised paper when finalizing their assessments, and we kindly request that they reconsider their scores in light of this evidence. We are glad to answer any further questions during the discussion period.

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

Related papers

Discussion (0)

Sign in to comment.