Multi-Scale Mamba (MS-Mamba) Research

• MS-Mamba is a family of state-space model architectures that integrate multiple scales (temporal, spatial, spectral) to enhance long-range dependency modeling.
• It leverages techniques like parallel scans, downsampling, and hierarchical pyramids to transform single-scale models into efficient, task-specific multi-scale systems.
• Empirical results show MS-Mamba variants achieve improved accuracy-efficiency trade-offs across various applications, including vision, remote sensing, and time-series forecasting.

The literature suggests that Multi-Scale Mamba (MS-Mamba) denotes a family of Mamba- or selective state-space-model architectures that introduce explicit multi-scale processing into otherwise single-scale sequence models. Across vision, remote sensing, time-series forecasting, reinforcement learning, recommendation, and temporal action detection, these systems retain the linear-complexity selective scan paradigm of Mamba while altering how inputs are partitioned, sampled, scanned, fused, or hierarchically organized. The common objective is to couple long-range dependency modeling with scale diversity—spatial, temporal, spectral, frequency-domain, or modality-level—without reverting to quadratic self-attention (Karadag et al., 10 Apr 2025, Shi et al., 2024).

1. Conceptual scope and terminology

Taken together, the published uses of the term indicate that “multi-scale” is not standardized in Mamba research. In some works it means multiple temporal sampling rates, in others full- and low-resolution spatial scans, in others a U-Net pyramid, dilated state-space transitions, or frequency-domain decomposition. The term therefore names a design pattern rather than a single canonical block.

Instantiation	Domain	Multi-scale realization
ms-Mamba (Karadag et al., 10 Apr 2025)	Time-series forecasting	Parallel Mamba blocks with different sampling rates $\Delta_i$
MSVMamba (Shi et al., 2024)	Vision backbone	One full-resolution scan plus three downsampled scans
MSFMamba (Gao et al., 2024)	Multi-source remote sensing	Full- and half-resolution spatial SSM routes inside MSpa-Mamba
M3SR (Zhang et al., 13 Jan 2026)	Spectral reconstruction	U-Net structure with global, intermediate, and local scales
MS-Temba (Sinha et al., 10 Jan 2025)	Temporal action detection	Block-wise dilated temporal convolution and dilated SSM
M2Rec (Zhang et al., 7 May 2025)	Sequential recommendation	Temporal Mamba, FFT-based frequency modeling, and semantic fusion

Related medical-imaging variants reinforce this plurality. MSV-Mamba introduces a decoder-side large-window multiscale Mamba module with pixel-level and patch-level branches (Yang et al., 13 Jan 2025). MSVM-UNet inserts multi-scale convolutions into VSS blocks and combines them with large-kernel patch expanding (Chen et al., 2024). A 3D volumetric study defines a multi-scale Mamba block, denoted MSv4, around parallel $3\times3\times3$, $5\times5\times5$, and $7\times7\times7$ depthwise convolutions followed by Mamba scanning (Wang et al., 25 Mar 2025).

2. State-space foundations and how scale is introduced

Most MS-Mamba variants preserve the standard state-space backbone. A continuous-time SSM is typically written as

$$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t), \qquad y(t)=\mathbf{C}h(t),$$

with discretization by a step size $\Delta$ yielding

$$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t.$$

Mamba then makes key quantities such as $\mathbf{B}$, $\mathbf{C}$, and $\Delta$ input-dependent, producing a selective SSM with linear complexity in sequence length (Karadag et al., 10 Apr 2025).

What changes in MS-Mamba is usually not the recurrence itself, but the way sequences are formed or parameterized around it. In ms-Mamba for forecasting, several Mamba blocks run in parallel with different sampling rates $3\times3\times3$0, and their outputs are averaged, so the same input is processed at several temporal scales (Karadag et al., 10 Apr 2025). In Multi-Scale VMamba, a feature map is split into a full-resolution branch and a downsampled branch; one scan is applied to the former and three scans to the latter, after which low-resolution outputs are interpolated and fused back (Shi et al., 2024). In MS-Temba, dilation is pushed into the SSM by a reshaping operator $3\times3\times3$1 that partitions a sequence into subsequences such as odd and even indices, applies BiMamba to each, and then restores the original layout through $3\times3\times3$2 (Sinha et al., 10 Jan 2025). In offline RL, MambaDM uses a parallel global–local fusion Mamba module in which one branch processes the full trajectory while another processes fixed-length windows, making multi-scale dependence explicit at the architecture level rather than inside a single scan kernel (Cao et al., 2024).

This suggests that MS-Mamba is best understood as scale-aware selective scanning: the SSM core remains intact, while scale enters through multiple $3\times3\times3$3s, dilated index mappings, multi-resolution feature maps, grouped channels, or explicit local/global branches.

3. Spatial, spectral, and hierarchical vision realizations

In image-like domains, MS-Mamba is frequently realized through multi-resolution scanning plus hierarchical pyramids. Multi-Scale VMamba is exemplary: it replaces four full-resolution SS2D scans with a “hierarchy-in-hierarchy” design that performs one scan at full resolution and three on a downsampled map, then supplements token mixing with a ConvFFN for channel mixing (Shi et al., 2024). The motivation is explicit: downsampling reduces sequence length, improves effective long-range modeling, and lowers cost, while a retained full-resolution scan preserves fine detail.

MSFMamba, developed for joint classification of hyperspectral image (HSI) and LiDAR/SAR data, gives a more task-specific but closely related formulation. The network stacks spatial–spectral Mamba modules, each applying MSpa-Mamba to both modalities, Spe-Mamba to HSI only, and Fus-Mamba for cross-modal fusion. Its multi-scale component is the Multi-Scale Spatial SSM inside MSpa-Mamba: depth-wise convolutions with strides 1 and 2 generate full- and half-resolution maps, two scanning routes operate at full resolution, two at half resolution, and the half-resolution outputs are upsampled before aggregation. The same model also separates spatial, spectral, and fusion roles across MSpa-Mamba, Spe-Mamba, and Fus-Mamba, respectively (Gao et al., 2024).

M3SR, for RGB-to-HSI spectral reconstruction, embeds Mamba in a U-Net-like encoder–decoder with three conceptual scales—global, intermediate, and local—and assigns a multi-perceptual fusion block to each scale. That block combines a spatial branch built on VMamba’s VSS/SS2D, a frequency branch built on DWT–VSS–IDWT, and a spectral branch that treats per-pixel spectra as 1D sequences for Mamba processing. Here the multi-scale notion is inseparable from the encoder–decoder pyramid, while the multi-perceptual notion adds domain transforms inside each scale (Zhang et al., 13 Jan 2026).

Medical imaging variants extend the same logic. MSV-Mamba uses a U-shaped model whose decoder contains an LMS module composed of PiM and PaM: PiM performs window-level pixel-sequence Mamba, whereas PaM pools each window to a token, applies bidirectional Mamba across tokens, and unpools back to full resolution. MSAA then fuses decoder features with spatial and channel attention, and auxiliary losses supervise multiple decoder depths (Yang et al., 13 Jan 2025). MSVM-UNet instead augments VSS blocks with an MS-FFN that sums depth-wise convolutions of sizes $3\times3\times3$4, $3\times3\times3$5, and $3\times3\times3$6, and replaces ordinary patch expanding with large-kernel patch expanding so that spatial and channel information are jointly integrated during upsampling (Chen et al., 2024). In 3D segmentation, the UlikeMamba study inserts multi-scale DWConv3D branches before Mamba and analyzes scan strategies ranging from single-scan to Tri-scan, thereby turning scale selection and scan complexity into explicit design variables (Wang et al., 25 Mar 2025).

4. Temporal, decision, and recommendation variants

In sequential domains, MS-Mamba is usually defined over temporal scales rather than spatial resolutions. ms-Mamba for multivariate forecasting is the most literal instance: each layer runs several bidirectional Mamba blocks with distinct sampling rates $3\times3\times3$7, and the outputs are averaged before normalization and an MLP. The paper evaluates fixed, learnable, and dynamic temporal scales, with the fixed multiplier set $3\times3\times3$8 and learnable $3\times3\times3$9 both performing strongly (Karadag et al., 10 Apr 2025).

MambaDM for offline reinforcement learning uses a different temporal interpretation. Its GLoMa layer splits a reward-conditioned trajectory into a global branch over the full sequence and a local branch over fixed-length windows, applies Mamba to both, fuses them additively, and follows with an FFN. The motivation is specific to RL trajectories, which exhibit both local Markovian structure and global trajectory-scale correlations (Cao et al., 2024).

MS-Temba adopts a block-wise temporal hierarchy for densely labeled temporal action detection. Each Temba Block contains a Temporal Local Module, implemented as a 1D temporal convolution with dilation $5\times5\times5$0, followed by a Dilated Temporal SSM in which the sequence is rearranged by $5\times5\times5$1 into stride-$5\times5\times5$2 subsequences, processed by BiMamba, and restored to the original order. Outputs from several such blocks are projected to a common dimension and fused by a Temporal Mamba Fuser, which itself uses BiMamba after summation (Sinha et al., 10 Jan 2025).

M2Rec shows that some authors extend “multi-scale” beyond time alone. Its Adaptive FFT–Powered Mamba encoder first transforms a user interaction sequence to the frequency domain, masks high-frequency components, reconstructs a denoised sequence by inverse FFT, and then applies Mamba. The resulting temporal-frequency representation is fused with LLM-based semantic item embeddings through a learnable gate. In this usage, the scales are temporal, frequency-domain, and semantic rather than multiple SSM kernels in the narrow sense (Zhang et al., 7 May 2025).

5. Empirical behavior, efficiency, and representative results

The empirical record consistently links MS-Mamba designs to improved accuracy-efficiency trade-offs, though the exact benefit depends on domain and realization. On ImageNet, COCO, and ADE20K, MSVMamba-Tiny reports 82.8% top-1 accuracy, 46.9% box mAP, 42.2% instance mAP, and 47.6% mIoU with single-scale testing on ADE20K (Shi et al., 2024). In time-series forecasting, ms-Mamba on ETTh2 at horizon 96 achieves MSE 0.291, params 0.481M, memory 1.84MB, and MACs 0.165G, compared with MSE 0.296, params 1.150M, memory 4.40MB, and MACs 1.563G for S-Mamba (Karadag et al., 10 Apr 2025).

In multi-source remote sensing, MSFMamba reports the best overall accuracy and Kappa on Berlin, Augsburg, and Houston2018, with OA 77.11% / 91.30% / 92.38% and, on Augsburg, 1.5252M parameters and 0.0377G FLOPs, compared with 37.7M parameters and 3.56G FLOPs for FusAtNet (Gao et al., 2024). In spectral reconstruction, M3SR on NTIRE2022 reaches RMSE 0.0343, PSNR 31.3995 dB, SAM 6.6199, MSSIM 0.9351, with 2.166M parameters and 100.924G FLOPs for $5\times5\times5$3 input (Zhang et al., 13 Jan 2026). In long-video temporal action detection, MS-Temba reports 10M temporal parameters and 0.6 GFLOPs, versus 87M and 6.6 GFLOPs for MS-TCT, while remaining state-of-the-art on long-duration videos (Sinha et al., 10 Jan 2025). In 3D volumetric segmentation, the integrated UlikeMamba_3dMT reports Dice 89.95 on AMOS and 90.60 on BraTS, with 93.09G FLOPs (Wang et al., 25 Mar 2025).

These results do not imply a single dominant recipe. Rather, they suggest that explicit scale handling—through downsampled scans, multiple $5\times5\times5$4s, dilated SSMs, or hierarchical feature pyramids—can convert Mamba’s linear-time recurrence into a more task-aligned inductive bias.

6. Misconceptions, limitations, and open directions

A common misconception is that “multi-scale” always means multi-resolution images. The literature does not support that reading. In MSFMamba it is full- and half-resolution spatial SSM routes; in ms-Mamba it is multiple temporal sampling rates $5\times5\times5$5; in M2Rec it includes frequency-domain and semantic scales in addition to time; and in MambaDM it is the coexistence of local windows and full trajectories (Gao et al., 2024, Karadag et al., 10 Apr 2025, Zhang et al., 7 May 2025, Cao et al., 2024).

A second misconception is that more elaborate scanning or larger models are automatically beneficial. The 3D volumetric study explicitly finds that simpler methods often suffice, while Tri-scan yields notable advantages only in the most challenging scenarios (Wang et al., 25 Mar 2025). MambaDM reports a scaling-law result in which increasing model size does not bring performance improvement, whereas scaling the dataset amount by 2x for MambaDM can obtain up to 33.7% score improvement on Atari dataset (Cao et al., 2024). MSVM-UNet shows that the kernel set $5\times5\times5$6 outperforms $5\times5\times5$7 and $5\times5\times5$8, indicating that larger or more numerous branches can degrade HD95 and increase compute (Chen et al., 2024). MS-Temba similarly finds that three Temba blocks are optimal, while a fourth sharply reduces TSU mAP (Sinha et al., 10 Jan 2025).

Open directions in the literature are correspondingly pragmatic. MSFMamba explicitly mentions hybrid Mamba + Transformer frameworks and improved interpretability for multi-source remote sensing (Gao et al., 2024). M3SR also points to hybridization and suggests that the current L1-only objective could be extended by other losses, though it does not implement them (Zhang et al., 13 Jan 2026). This suggests that the next phase of MS-Mamba research is likely to concern how scale should be allocated, fused, and interpreted, rather than whether selective SSMs can replace attention in a uniform way.