acceptodds
Under review as a conference paper at ICLR 2027

LoopMamba: Rethinking State Space Models from an Energy Perspective

Abstract

As a powerful alternative to Transformers, State Space Models (SSMs) have demonstrated remarkable effectiveness across a wide range of vision tasks. However, existing visual SSM architectures still rely substantially on empirically motivated design choices to adapt SSMs to visual tasks, with limited theoretical understanding of their underlying state transition mechanisms to guide architectural design. In this work, we rethink SSMs from an energy perspective, revealing that state transitions can be viewed as an iterative energy minimization process in which hidden states progressively move toward stable representations. Guided by this insight, we propose LoopMamba, a visual architecture with four Self-loop Layers that performs iterative hidden state refinement through Variable-Parameter State Transition. Extensive experiments and analyses on image classification, object detection, instance segmentation, and semantic segmentation demonstrate the effectiveness of LoopMamba. On ImageNet-1K, LoopMamba achieves a 1.8% improvement in Top-1 accuracy over Vim-S with 31.5% fewer parameters. Overall, LoopMamba offers an effective and interpretable paradigm for SSM-based visual representation learning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
State Space Models, Energy-Based Models, Neural Network Architecture, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qTR4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that the SSM state update ht=Aˉht−1+Bˉxth_t=\bar A h_{t-1}+\bar B x_t can be read as one gradient-descent step on the quadratic energy E(x,h)=12ηh⊤(I−Aˉ)h−1η(Bˉx)⊤hE(x,h)=\tfrac{1}{2\eta}h^\top(I-\bar A)h-\tfrac{1}{\eta}(\bar B x)^\top h, and concludes from this that stacking SSM layers amounts to successive optimisation steps. On that basis it proposes LoopMamba, a four-stage hierarchical backbone with a single block per stage, in which a weight-shared Self-loop Layer is applied four or eight times. Each iteration also receives an Iteration-specific Lightweight State Fusion module, which combines three dilated depth-wise convolutions over the hidden state with a per-iteration low-rank adapter. On ImageNet-1K, LoopMamba-T-4 and T-8 reach 79.7% and 80.2% with 8.1M and 8.5M parameters, and LoopMamba-B-4 and B-8 reach 82.1% and 82.4% with 17.8M and 18.6M parameters at 6.4 and 9.6 GFLOPs. The paper also reports Mask R-CNN 1× results on COCO, UperNet results on ADE20K, ablations of the fusion components and of loop depth, an analysis showing that a hidden-state entropy decreases with loop depth, which it offers as evidence for the energy view, and qualitative feature maps compared with VMamba.

Critical strengths

The design is simple and parameter-efficient: with one block per stage and shared weights across loop iterations, LoopMamba reaches 82.1–82.4% top-1 with fewer than 19M parameters, and the depth sweep in Table 5 shows that additional iterations buy accuracy with almost no growth in parameters. The component ablations in Tables 4 and 5 are also informative and candid, since they show that most of the gain comes from the per-iteration low-rank adapters rather than from repeating the shared transition.

Critical weaknesses

1. Neither the looped architecture nor its energy-based motivation is new or specific to the design. Sharing one block across depth while adding lightweight, non-shared parameters at each iteration is the core design of SReT [1] and MiniViT [2], which are neither compared nor discussed. SReT already reports the finding that LoopMamba presents in Table 5, that naive recursion barely helps, from 72.2% to 72.5% in SReT and from 79.0% to 79.3% in LoopMamba, and that the gain appears only with non-shared per-iteration parameters. SReT-S also reaches 81.9% with 20.9M parameters and 4.2 GFLOPs, on par with LoopMamba-B-4 at 82.1% and 6.4 GFLOPs. The theoretical motivation also does not fill this gap. Theorem 1 rewrites a symmetric affine update as gradient descent on the quadratic in Eq.5, but the identity holds for any affine map with a symmetric linear part, and for Mamba's diagonal Aˉ\bar A it simply says that each coordinate's exponential moving average is a gradient step on a one-dimensional parabola. More importantly, it concerns the update along the sequence inside one scan, whereas LoopMamba loops over depth through convolutions, adapters, normalisation and an FFN, for which no energy is defined. the entropy plot in Fig. 4 is undefined, uses truncated per-subplot axes and has no non-looped control.

2. The model is not competitive at matched compute, and its efficiency is never measured. The paper claims "the best performance" and a "favorable accuracy–efficiency trade-off", but looping converts saved parameters into extra FLOPs and sequential depth. VMamba-T, which the paper uses in Fig. 5 but omits from Table 1, reaches 82.6% at 4.9 GFLOPs, ahead of LoopMamba-B-8 at 82.4% with 9.6 GFLOPs, and also leads on COCO and ADE20K with about 47.3 box AP, 42.7 mask AP and 47.9 mIoU against 46.1, 41.9 and 45.8. VSSD-Micro [3] reaches 82.5% with 14M parameters and 2.3 GFLOPs, beating LoopMamba-B-8 on accuracy, parameters and FLOPs at once. The training recipe also adds MESA and EMA, which several baselines did not use. There is no measured latency, throughput or memory, although inference latency grows with loop depth and training memory with the number of stored activations, and Tables 2 and 3 list the classification-backbone parameters and FLOPs for the detectors and segmenters, hiding exactly the resolution-dependent cost that matters for a looped backbone.

References

[1] Z. Shen, Z. Liu, E. Xing. Sliced Recursive Transformer. ECCV, 2022.

[2] J. Zhang, H. Peng, K. Wu, M. Liu, B. Xiao, J. Fu, L. Yuan. MiniViT: Compressing Vision Transformers with Weight Multiplexing. CVPR, 2022.

[3] Y. Shi, M. Dong, M. Li, C. Xu. VSSD: Vision Mamba with Non-Causal State Space Duality. arXiv:2407.18559, 2024.

Questions for authors
  1. How does LoopMamba relate to SReT and MiniViT, and what does the SSM setting add beyond them? Please also report an unshared stack with the same number of layers per stage at the same FLOPs, trained with your exact recipe.

  2. What are the throughput, latency and peak memory of LoopMamba-T-4, T-8, B-4 and B-8 compared with VMamba-T, VSSD-Micro or similar SoTA models on the same GPU?

12 comments
AuthorsOpenReview ↗

Response to reviewer's comments: Relation to SReT and MiniViT

We acknowledge the relevance of SReT and MiniViT to our work. However, unlike these Transformer-based approaches, LoopMamba focuses on the state transition mechanism of SSMs. Specifically, motivated by an energy-based interpretation of SSM state updates, we introduce VPST, which combines shared state transitions with iteration-specific state fusion. MSSF and ISLA further enable multi-scale hidden-state aggregation and iteration-specific low-rank adaptation, respectively. These designs target the iterative refinement of SSM hidden states under an interpretable view, rather than simply applying conventional parameter sharing to another backbone. We now include the discussion of SReT and MiniViT in Appendix I.

AuthorsOpenReview ↗

Response to reviewer's comments: Regarding naive recursion in SReT and LoopMamba

We would like to clarify the reviewer's interpretation of the naive recursion experimental results. The improvement from 72.2% to 72.5% cited by the reviewer comes from Table 3 of SReT, where the baseline and the recursive model without NLL have 5.7M and 3.8M parameters, respectively. However, Table 1 of SReT provides a more direct comparison, showing that one recursion improves DeiT-Tiny's Top-1 accuracy from 72.2% to 74.0% (+1.8 percentage points) while maintaining 5.7M parameters. Therefore, characterizing the benefit of naive recursion as merely 0.3 percentage points does not fully reflect the reported results. Similarly, Table 5 of LoopMamba shows that increasing the loop depth from 4 to 8 without ISLA improves accuracy from 79.0% to 79.3% without increasing parameters. These observations suggest that recurrent depth can benefit representation learning even without additional iteration-specific adaptation, although the improvement may be limited by the capacity of shared transformations.

AuthorsOpenReview ↗

Response to reviewer's comments: Regarding the comparison with SReT-S

We believe that accuracy, parameter count, and computational complexity should be considered jointly. SReT-S achieves 81.9% Top-1 accuracy with 20.9M parameters and 4.2G FLOPs, whereas LoopMamba-B-4 achieves 82.1% with 17.8M parameters and 6.4G FLOPs. Although LoopMamba-B-4 requires more computation, it reduces the parameter count by approximately 14.8% while achieving slightly higher accuracy. These results highlight different efficiency trade-offs: SReT offers lower computational complexity, whereas LoopMamba demonstrates better parameter efficiency at this model scale.

Reviewer_8L2nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LoopMamba, a visual state space model architecture motivated by an energy-based interpretation of SSM state transitions. The authors show that, under certain assumptions, a discrete SSM state transition can be expressed as a gradient descent step on a quadratic energy function. Based on this interpretation, LoopMamba introduces Variable-Parameter State Transition (VPST), which repeatedly applies shared state transition parameters for iterative hidden state refinement. An Iteration-specific Lightweight State Fusion (ILSF) module further incorporates multi-step state fusion and iteration-specific low-rank adaptation to improve representation flexibility.

Experiments on ImageNet-1K, COCO, and ADE20K demonstrate competitive performance in image classification, object detection, instance segmentation, and semantic segmentation. The results suggest that LoopMamba can achieve favorable parameter efficiency while maintaining strong visual representation capabilities.

Critical strengths

The paper attempts to connect SSM state transitions with energy-based optimization and uses this perspective to motivate iterative hidden state refinement. This provides an interesting interpretation of state transitions and a clear motivation for the proposed architecture.

LoopMamba achieves strong results across image classification, object detection, instance segmentation, and semantic segmentation. In particular, LoopMamba-B-4 improves ImageNet-1K Top-1 accuracy by 1.8 percentage points over Vim-S while using approximately 31.5% fewer parameters. The evaluation across multiple downstream tasks supports the practical effectiveness of the proposed architecture.

Critical weaknesses

Theorem 1 establishes an equivalence between a linear SSM state transition and a gradient descent step on a quadratic energy function under specific assumptions. However, the actual LoopMamba architecture incorporates nonlinear transformations, normalization, residual connections, and iteration-specific state fusion. It remains unclear whether the complete architecture preserves the proposed energy minimization interpretation.

Furthermore, the theoretical formulation concerns state transitions across sequence positions, whereas LoopMamba performs repeated refinement across network iterations. The relationship between these two processes requires further justification. Although Figure 4 shows decreasing hidden-state entropy, entropy reduction does not necessarily imply energy minimization. Consequently, the current evidence does not sufficiently support the central theoretical motivation.

The proposed self-loop mechanism is closely related to existing parameter-sharing and recurrent-depth architectures. Although iteration-specific low-rank adaptation improves flexibility, the paper lacks direct comparisons with weight-tied Mamba architectures equipped with similar adaptation mechanisms. Such comparisons are important for establishing whether VPST provides benefits beyond conventional parameter sharing.

Moreover, increasing loop depth substantially increases computational cost. For example, increasing the loop depth from 4 to 8 improves ImageNet accuracy from 79.7% to 80.2%, while FLOPs increase from 2.9G to 4.4G. The reported results demonstrate parameter efficiency, but the practical computational efficiency remains unclear without inference latency and throughput measurements.

Questions for authors

Does the complete LoopMamba architecture, including ILSF and other nonlinear components, correspond to the minimization of a well-defined energy function? If so, can the authors provide theoretical justification or direct empirical evidence of energy reduction across iterations? This would substantially strengthen the paper's central claim.

How does LoopMamba compare with a weight-tied Mamba architecture incorporating iteration-specific low-rank adaptation under matched parameter and FLOPs budgets? This comparison would help clarify the contribution of VPST beyond existing recurrent-depth and parameter-sharing designs.

Can the authors provide actual inference latency or throughput measurements for different loop depths? Given the increasing FLOPs, these measurements would help determine whether the proposed architecture offers practical efficiency advantages beyond parameter reduction.

6 comments
AuthorsOpenReview ↗

Response to reviewer's comments: Concern about Theorem 1

We thank the reviewer for highlighting these two important questions.

We have added analyses in Appendix B and Appendix C to address the full recurrence and its relationship to sequence-wise state transitions.

Regarding the energy interpretation of the full recurrence, Theorem 1 establishes the equivalence for a single conditional SSM transition.

Our additional analysis explicitly expands the complete repeated update zk=Fk(zk−1)z^k=F_k(z^{k-1}), including the SSM scan, MSSF, ISLA, normalization, gating, and skip connection.

Following the implemented architecture, the outer residual connection and FFN are applied outside the repeated Self-loop Layer.

With trained parameters fixed, we define a reference map Fˉ=∑j=1KωjFj\bar F=\sum_{j=1}^{K}\omega_jF_j, where the positive averaging weights sum to one and remain fixed.

For continuously differentiable maps on an open convex domain containing a fixed reference ζ\zeta, we construct

E(z)=1η∫01⟨ξs−Fˉ(ξs),z−ζ⟩ds,ξs=ζ+s(z−ζ).\mathcal E(z)=\frac{1}{\eta}\int_0^1\left\langle \xi_s-\bar F(\xi_s),z-\zeta\right\rangle ds,\qquad \xi_s=\zeta+s(z-\zeta).

This is a common energy across iterations, with changing representations treated as its state arguments.

Proposition 1 establishes the exact decomposition:

zk=zk−1−η∇E(zk−1)+ek(zk−1),ek=Fk−Fˉ+ηa,z^k=z^{k-1}-\eta\nabla\mathcal E(z^{k-1})+e_k(z^{k-1}),\qquad e_k=F_k-\bar F+\eta a,

where aa is determined by the antisymmetric part of the reference Jacobian.

The two correction terms account for iteration-specific adaptation and nonconservative transformations.

The appendix expands the component Jacobians and derives bounds on these corrections.

Exact gradient descent holds when ek=0e_k=0.

For nonzero corrections, Theorem 2 establishes energy decrease and finite-depth bounds when the iterates remain in a convex region where the energy is lower bounded, its gradient is LL-Lipschitz, and

∥ek(zk−1)∥≤αη∥∇E(zk−1)∥,0≤α<1,ηL<2(1−α)(1+α)2.\|e_k(z^{k-1})\|\le\alpha\eta\|\nabla\mathcal E(z^{k-1})\|,\qquad 0\le\alpha<1,\qquad \eta L<\frac{2(1-\alpha)}{(1+\alpha)^2}.

These assumptions accommodate nonlinear operations and iteration-specific parameters through their contributions to the complete map and correction bounds.

They are sufficient conditions that must be verified or enforced for the trained maps.

For the implemented 4- or 8-loop networks, the applicable results are finite-depth bounds.

An infinite extension satisfying the same assumptions yields vanishing gradients, with state convergence additionally established under strong convexity.

Regarding sequence positions and network iterations, Appendix C explicitly separates the sequence index tt from the iteration index kk.

Within iteration kk, the scan receives uk−1=u(zk−1)u^{k-1}=u(z^{k-1}) and computes

htk=Aˉt(uk−1)ht−1k+Bˉt(uk−1)utk−1.h_t^k=\bar A_t(u^{k-1})h_{t-1}^k+\bar B_t(u^{k-1})u_t^{k-1}.

Under the assumptions of Theorem 1, each such transition has a conditional quadratic-energy interpretation.

Unrolling these transitions produces the complete hidden-state representation HkH^k.

State fusion, readout, normalization, and gating then transform HkH^k into zkz^k, which generates the next scan input uk=u(zk)u^k=u(z^k).

Thus, sequence-wise aggregation in one iteration influences the driving features and selective coefficients of the next iteration.

The added analysis derives both the unrolled scan and an exact identity for hidden-state differences between successive iterations, making this feedback relationship explicit.

The local conditional energies describe the transitions inside each scan, while the fixed energy E\mathcal E describes the composite recurrence through the correction decomposition above.

Together, these results connect the single-transition interpretation to the iterative architecture and specify the additional conditions required for energy decrease across network iterations.

AuthorsOpenReview ↗

Response to reviewer's comments: Concern about energy minimization

To directly assess energy evolution, we now evaluate intermediate representations across iterations and LoopMamba variants using the fixed stage-specific energy functions defined in Appendix B. The results in Sec. 4.6 and Appendix H show net energy reductions across all stages, with occasional intermediate increases. For example, in LoopMamba-T-4, Stage 1 exhibits a monotonic decrease from 0.0336 to -0.0345 over four iterations, while Stage 4 decreases overall from -0.0897 to -0.2078. Similar trends are observed in the larger and deeper variants, with Stage 1 energy decreasing from 0.0697 to -0.0166 in LoopMamba-T-8 and from 0.1138 to -0.0370 in LoopMamba-B-8. These within-stage measurements provide direct empirical support for energy-based refinement and complement the entropy analysis. The observed local fluctuations are consistent with our theoretical formulation, which allows corrections arising from nonconservative transformations and iteration-specific adaptation. Monotonic descent and convergence remain conditional on the assumptions stated in Appendix B.

AuthorsOpenReview ↗

Response to reviewer's comments: Compare with weight-tied Mamba

We thank you for your constructive comment. To address this comparison, we have added an ISLA-only ablation that removes MSSF while retaining the shared SSM transitions and iteration-specific low-rank adapters. This variant provides a controlled weight-tied Mamba baseline with iteration-specific adaptation under closely matched parameter and FLOPs budgets.

Table-4: Ablation study of Variable-Parameter State Transition on ImageNet-1K. MSSF: multi-step state fusion. ISLA: iteration-specific low-rank adaptation.

LoopMamba-T-4

MSSFISLAP. (M)F. (G)Top-1 (%)
××7.72.7878.8
✓×7.82.8679.0 (+0.2)
×✓8.12.8379.4 (+0.6)
✓✓8.12.9179.7 (+0.9)
Unshared stack—23.34.1781.3

As shown in Table 4, it achieves 79.4% Top-1 accuracy with 8.1M parameters and 2.83G FLOPs, compared with 79.7%, 8.1M parameters, and 2.91G FLOPs for the full LoopMamba-T-4. Thus, incorporating MSSF improves accuracy by 0.3 percentage points with the same reported parameter count and approximately 2.8% additional FLOPs. The ISLA-only variant also improves upon the baseline without either component by 0.6 percentage points. Together, these results demonstrate the contribution of iteration-specific adaptation and the additional benefit of multi-step state fusion beyond shared transitions equipped with low-rank adapters.

Reviewer_f8k7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes LoopMamba, a visual state-space model that interprets a single SSM state transition as a gradient-descent step on a quadratic energy function under specified conditions. Based on this interpretation, the authors introduce Variable-Parameter State Transition (VPST), which reuses state-transition parameters across iterations, and Iteration-specific Lightweight State Fusion (ILSF), which combines multi-scale depth-wise convolutions with iteration-specific low-rank adaptation. Experiments on ImageNet-1K, COCO, and ADE20K show competitive results in image classification, object detection, instance segmentation, and semantic segmentation. The paper also analyzes model depth and hidden-state entropy.

Critical strengths
  1. The paper attempts to connect SSM state transitions with an energy-based optimization view and uses this perspective to motivate an iterative visual architecture. This offers a potentially useful way to reason about state-space model design.

  2. VPST shares state-transition parameters across iterations, while ILSF adds lightweight iteration-specific adaptation. This is a plausible way to increase effective computational depth without a proportional increase in parameter count. The evaluation covers several visual recognition tasks.

Critical weaknesses
  1. The theoretical analysis establishes an equivalence between a single SSM state transition and one gradient-descent step under stated assumptions. It does not establish that the full LoopMamba recurrence monotonically decreases one fixed energy function or converges. In VPST, the input representation changes across iterations, and ILSF further modifies the hidden state. The paper needs to clarify whether a common energy function governs the full iterative process and what conditions, if any, guarantee energy decrease or convergence. This is central to the paper’s energy-based motivation.

  2. The experiments do not yet provide sufficient evidence for several central claims. The decreasing hidden-state entropy reported in Section 4.6 does not by itself demonstrate energy minimization or convergence. The ablation in Table 4 also lacks a variant with ISLA alone, so the independent contribution of this component cannot be assessed. In addition, the semantic-segmentation results are mixed, and the strongest LoopMamba-B-8 result comes with 9.6G FLOPs. More direct state-evolution analysis, a more complete ablation, and comparisons under similar parameter or computation budgets would make the evidence more convincing.

Questions for authors
  1. Does the full VPST/ILSF iteration correspond to descent on a single, fixed energy function? If so, please state the required assumptions and explain how the changing input representations and iteration-specific parameters satisfy them. If not, please clarify the scope of the energy-based interpretation.

  2. How is hidden-state entropy computed in Section 4.6, and what evidence supports using its decrease as an indicator of refinement? Can the authors provide direct measurements of energy change or state convergence?

  3. Can the authors report an ISLA-only ablation and compare LoopMamba with recurrent or parameter-sharing baselines under comparable parameter counts or FLOPs?

4 comments
AuthorsOpenReview ↗

Response to reviewer's comments: Concern about theoretical analysis

We thank the reviewer for identifying this gap in our original analysis. To address the full recurrence, we have added Appendix B, which constructs a common energy for the complete self-loop update and establishes sufficient conditions for energy decrease and convergence.

Within a fixed-dimensional stage, we write the implemented recurrence as zk=Fk(zk−1)z^k=F_k(z^{k-1}), where FkF_k explicitly includes the SSM scan, MSSF, ISLA, normalization, gating, and feature feedback. With trained parameters fixed, we define Fˉ(z)=∑j=1KωjFj(z)\bar F(z)=\sum_{j=1}^{K}\omega_jF_j(z), using fixed positive weights satisfying ∑jωj=1\sum_j\omega_j=1. For continuously differentiable maps on an open convex domain containing a fixed reference ζ\zeta, we construct

E(z)=1η∫01⟨ξs−Fˉ(ξs),z−ζ⟩ds,ξs=ζ+s(z−ζ).\mathcal E(z)=\frac{1}{\eta}\int_0^1\left\langle \xi_s-\bar F(\xi_s),z-\zeta\right\rangle ds,\qquad \xi_s=\zeta+s(z-\zeta).

This energy remains fixed across iterations because its defining maps, weights, and reference remain fixed. The changing representations are its state arguments, and input-dependent SSM coefficients are retained within the maps FkF_k. For a single conditional affine SSM transition with ζ=0\zeta=0, this construction recovers the quadratic energy of Theorem 1.

Proposition 1 then gives the exact decomposition:

zk=zk−1−η∇E(zk−1)+ek(zk−1),ek(z)=Fk(z)−Fˉ(z)+ηa(z),z^k=z^{k-1}-\eta\nabla\mathcal E(z^{k-1})+e_k(z^{k-1}),\qquad e_k(z)=F_k(z)-\bar F(z)+\eta a(z),

where

a(z)=1η∫01s(JFˉ(ξs)−JFˉ(ξs)⊤)(z−ζ) ds.a(z)=\frac{1}{\eta}\int_0^1 s\bigl(J_{\bar F}(\xi_s)-J_{\bar F}(\xi_s)^\top\bigr)(z-\zeta)\,ds.

Thus, iteration-specific adaptation contributes through Fk−FˉF_k-\bar F, while the nonconservative contribution of the composite update is captured by aa. Appendix B expands the scan, fusion, normalization, and gating Jacobians and bounds the effects of adapter differences on these corrections. Exact gradient descent on the common energy holds when ek=0e_k=0.

For nonzero corrections, Theorem 2 establishes energy decrease when the iterates remain in a convex region where E\mathcal E is lower bounded, its gradient is LL-Lipschitz, and

∥ek(zk−1)∥≤αη∥∇E(zk−1)∥,0≤α<1,ηL<2(1−α)(1+α)2.\|e_k(z^{k-1})\|\le\alpha\eta\|\nabla\mathcal E(z^{k-1})\|,\qquad 0\le\alpha<1,\qquad \eta L<\frac{2(1-\alpha)}{(1+\alpha)^2}.

Under these conditions:

E(zk)≤E(zk−1)−cη∥∇E(zk−1)∥2,cη=η(1−α)−Lη22(1+α)2>0.\mathcal E(z^k)\le\mathcal E(z^{k-1})-c_\eta\|\nabla\mathcal E(z^{k-1})\|^2,\qquad c_\eta=\eta(1-\alpha)-\frac{L\eta^2}{2}(1+\alpha)^2>0.

This yields finite-depth bounds for the implemented 4- or 8-loop networks. An infinite extension satisfying the same conditions has vanishing gradients, with convergence of the state sequence additionally established under strong convexity.

These results explicitly accommodate changing representations and iteration-specific parameters through the full maps and their correction terms. Their applicability to trained models requires the stated regularity and correction bounds, which are not automatically ensured by parameter sharing or low-rank adaptation. We therefore state the full-recurrence interpretation as an exact energy-gradient decomposition with conditional descent guarantees, and have revised the manuscript to distinguish these guarantees from the single-transition equivalence.

AuthorsOpenReview ↗

Response to reviewer's comments: Concern about entropy computing and energy evaluation

We thank the reviewer for this constructive comment.

In Section 4.6, we compute the hidden-state entropy as follows (now included in Appendix D):

zb,i=1C∑c=1CXb,c,i,z_{b,i} = \frac{1}{C}\sum_{c=1}^{C} X_{b,c,i},

pb,i=exp⁡(zb,i)∑j=1Nexp⁡(zb,j),p_{b,i} = \frac{\exp(z_{b,i})}{\sum_{j=1}^{N}\exp(z_{b,j})},

Hb=−1log⁡N∑i=1Npb,ilog⁡(pb,i+ε)H_b = -\frac{1}{\log N}\sum_{i=1}^{N} p_{b,i}\log(p_{b,i}+\varepsilon)

where XX is the input, and ϵ=1e−8\epsilon = 1e-8, we compute entropy after the operation htk=Aˉht−1k+Bˉxtk−1h_t^k = \bar{\mathbf{A}}h_{t-1}^k + \bar{\mathbf{B}}x_t^{k-1} and before ILSF.

This implementation first averages the features in the channel, then calculates Softmax along the spatial dimension, and finally calculates normalized entropy.

The rationale is that, at a fixed spatial resolution, a broadly distributed response produces higher entropy, whereas a response concentrated on fewer positions produces lower entropy.

The observed entropy decrease provides descriptive evidence of increasing spatial selectivity across iterations.

This observation is consistent with the feature visualizations in Fig.6, which show that LoopMamba concentrates responses on the main semantic objects while suppressing irrelevant background regions.

To further investigate state evolution, we evaluate intermediate representations across all LoopMamba variants using the fixed stage-specific energy functions defined in Appendix B.

The results in Sec. 4.6 and Appendix H show net energy reductions across all stages, with occasional increases at intermediate iterations.

For example, Stage 1 energy decreases from 0.0336 to -0.0345 in LoopMamba-T-4, from 0.0697 to -0.0166 in LoopMamba-T-8, and from 0.1138 to -0.0370 in LoopMamba-B-8.

Each trajectory is evaluated against the same energy function throughout its iterations, providing direct measurements of energy change beyond the descriptive entropy statistic.

These observations provide empirical support for energy-based refinement over the implemented finite-depth recurrence, while monotonic descent and convergence remain conditional on the assumptions stated in Appendix B.

AuthorsOpenReview ↗

Response to reviewer's comments: Concern about ablation study and baseline comparison

We thank the reviewer for this constructive comment. To assess the contribution of ISLA, we have added an ISLA-only variant to the ablation study in Table 4.

LoopMamba-T-4

MSSFISLAP. (M)F. (G)Top-1 (%)
××7.72.7878.8
✓×7.82.8679.0 (+0.2)
×✓8.12.8379.4 (+0.6)
✓✓8.12.9179.7 (+0.9)
Unshared stack—23.34.1781.3

Adding ISLA alone improves ImageNet-1K Top-1 accuracy from 78.8% to 79.4%, a gain of 0.6 percentage points, while increasing parameters from 7.7M to 8.1M and FLOPs from 2.78G to 2.83G. When combined with MSSF, ISLA improves accuracy from 79.0% to 79.7%, demonstrating its effectiveness both independently and within the complete ILSF module. Furthermore, the complete model exceeds the ISLA-only variant by 0.3 percentage points, indicating that MSSF provides a complementary benefit. These results support the contribution of iteration-specific low-rank adaptation to representation learning with modest parameter and computational overhead. Moreover, we have added an unshared-stack variant of LoopMamba, which achieves 81.3% Top-1 accuracy with 23.3M parameters and 4.17G FLOPs. Compared with this variant, LoopMamba-T-4 uses 65.2% fewer parameters and 30.2% fewer FLOPs, with a 1.6-percentage-point decrease in accuracy. At a similar computational budget, LoopMamba-T-8 achieves 80.2% accuracy with 8.5M parameters and 4.4G FLOPs, reducing the parameter count by 63.5% with a 1.1-percentage-point decrease in accuracy. Meanwhile, LoopMamba-B-4 achieves higher accuracy at 82.1% with fewer parameters (17.8M), although it requires more computation (6.4G FLOPs). These comparisons demonstrate the parameter efficiency of LoopMamba while clarifying the trade-offs between parameter sharing, computational cost, and accuracy.

Regarding semantic segmentation, as discussed in the paper, the recurrent architecture introduces additional computation while reducing the parameter count. LoopMamba uses fewer parameters than the compared Swin, Vim, and FastVim variants, and our baselines were selected primarily to provide approximately comparable parameter budgets. For a comparison under similar computational budgets in semantic segmentation, Swin-S achieves 47.6% mIoU with 50M parameters and 8.7G FLOPs, while LoopMamba-B-8 achieves 45.8% mIoU with 18.6M parameters and 9.6G FLOPs. Although their computational costs are comparable, Swin-S has approximately 2.7 times as many parameters as LoopMamba-B-8. In other words, LoopMamba-B-8 reduces the parameter count by approximately 62.8%, with a 1.8 percentage point decrease in mIoU. These results illustrate the trade-off between parameter efficiency, computational cost, and segmentation accuracy in our recurrent architecture. For a comparison with recurrent or parameter-sharing baselines under comparable parameter counts or FLOPs, because there are no other recurrent or parameter-sharing visual mamba baselines. To address this, we refer to the ablation study, without MSSF and ISLA, LoopMamba mirrors the Vim scanning mechanism, reflecting the recurrent version of Vim and reaching 78.8% on ImageNet, while the full LoopMamba (with MSSF and ISLA) outperforms it at 79.7% under comparable parameter size and computational budget.

Dear Reviewer f8k7,

As the discussion period nears its end, we hope that we have effectively addressed and resolved your concerns. Your feedback on our rebuttal responses would be greatly appreciated, and we are more than willing to provide further clarification on any remaining issues. Thank you for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary for Rebuttal

We sincerely thank all the reviewers and the Area Chair for their time, careful evaluation, and constructive feedback. We greatly appreciate the insightful comments, which have helped us identify important gaps in our original presentation and substantially improve the technical rigor and clarity of this work.

In response to the reviewers' concerns, we have made substantial revisions to the manuscript, with the major improvements summarized as follows:

  1. Strengthened theoretical analysis. We have extended the energy analysis from a single SSM transition to the complete Self-loop recurrence, establishing an exact gradient-plus-correction decomposition and deriving sufficient conditions for energy decrease, finite-depth guarantees, and conditional convergence. We have also clarified the relationship between sequence-wise state transitions and network-level iterations through an explicit analysis of feature feedback and cross-iteration state evolution. Furthermore, we provide visualizations of energy evolution across iterations, offering additional empirical support for our analysis.

  2. Expanded ablation studies and architectural comparisons. We have added direct energy evaluations across iterations and stages, an ISLA-only ablation, and an unshared-stack baseline. In addition, we have expanded the discussion of SReT and MiniViT and clarified the relationship between LoopMamba and existing recurrent-depth architectures.

  3. Improved efficiency evaluation and baseline comparisons. We have added comprehensive measurements of inference throughput, latency, and peak memory on the same GPU, providing a more practical assessment of different loop depths and representative SSM baselines.

  4. Corrected complexity reporting. We have recalculated the parameters and FLOPs of the complete detection and segmentation models, including both backbones and task-specific components. The corresponding tables and analyses have been updated to provide a more accurate and comprehensive assessment of downstream performance and computational costs.

We believe these revisions provide a substantially more rigorous and comprehensive account of LoopMamba, addressing the central concerns regarding its theoretical motivation, architectural contributions, empirical evidence, and practical efficiency. More importantly, they clarify that our contribution is not simply parameter sharing across network depth, but an SSM-oriented iterative refinement framework that connects conditional state-transition dynamics with network-level feature refinement through VPST and ILSF.

We are grateful for the opportunity to improve our work through this constructive review process. We respectfully ask the reviewers and the Area Chair to reconsider the revised manuscript in light of the additional theoretical results, experiments, and clarifications. We sincerely hope that the substantially strengthened analysis and evaluation demonstrate the value of LoopMamba and its potential contribution to energy-inspired visual state-space modeling. Thank you again for your time and consideration.

Area_Chair_6dHEOpenReview ↗

Reviewer-author discussion period is ending. Please feedback.

Dear Reviewers,

Thanks for your time and effort in reviewing ICLR2027 submissions.

As the reviewer-author discussion period will end very soon, please take time to read the author’s rebuttal and give your feedback. Your continuous engagement is highly appreciated.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.