FaRMamba: Hybrid Mamba Architectures

• FaRMamba is a shared label for multiple Mamba-based frameworks that blend state-space modeling with auxiliary modules like frequency transforms or fast attention to enhance domain-specific tasks.
• In physical field reconstruction, FR-Mamba fuses Fourier Neural Operators with linear-time state-space models to accurately reconstruct dynamic fields from sparse sensor data.
• For medical segmentation, FaRMamba integrates multi-scale frequency transforms and self-supervised reconstruction to restore high-frequency details and improve segmentation metrics.

FaRMamba is not a single canonical architecture in the recent arXiv record. In one usage, FR-Mamba is “also referred to as FaRMamba” and denotes a hybrid spatiotemporal operator-learning framework for Physical Field Reconstruction (PFR) from sparse measurements (Long et al., 21 May 2025). In another, FaRMamba is the formal name of a frequency-aware, reconstruction-aided Vision Mamba architecture for medical image segmentation (Rong et al., 26 Jul 2025). Related papers further state that the query term “FaRMamba” does not appear in the traffic image restoration model “Frequency-Aware Mamba,” is a colloquial match to FMamba for multivariate time-series forecasting, and refers to the Frequency-Assisted Mamba framework for remote sensing image super-resolution (Pan et al., 3 Dec 2025, Ma et al., 2024, Xiao et al., 2024). This suggests that the label functions as an overloaded name for several Mamba-centered designs rather than as a uniquely standardized model family.

1. Nomenclature and scope

The literature uses closely related names for distinct architectures, tasks, and inductive biases. The resulting ambiguity is substantive rather than typographic, because the same surface form is tied to different technical programs.

Context	Official name in paper	Relation to “FaRMamba”
Physical field reconstruction	FR-Mamba	“also referred to as FaRMamba”
Medical image segmentation	FaRMamba	official model name
Traffic image restoration	Frequency-Aware Mamba (FA-Mamba)	query term “FaRMamba” does not appear; “most likely a typo or informal variant”
Multivariate time-series forecasting	FMamba	“FaRMamba” colloquially captures “Fast-attention + Mamba”
Remote sensing image super-resolution	FMSR	“FaRMamba refers to” the Frequency-Assisted Mamba framework

The shared lexical core is the Mamba or state-space-model lineage, but the surrounding prefixes point to different emphases: field reconstruction, frequency awareness, fast attention, or frequency assistance. The technical content therefore has to be interpreted paper by paper rather than by name alone (Long et al., 21 May 2025, Rong et al., 26 Jul 2025, Pan et al., 3 Dec 2025, Ma et al., 2024, Xiao et al., 2024).

2. FR-Mamba / FaRMamba for physical field reconstruction

In "FR-Mamba: Time-Series Physical Field Reconstruction Based on State Space Model" (Long et al., 21 May 2025), FaRMamba designates a hybrid spatiotemporal operator-learning framework for reconstructing transient physical fields from sparse sensor observations. The target problem is Physical Field Reconstruction, where sparse measurements are used to infer time-varying distributions of quantities such as velocity, pressure, temperature, or vorticity. The specific benchmark is a canonical 2D unsteady flow past a cylinder, with transient vorticity field $u^t \in \mathbb{R}^{H \times W}$ as the reconstruction target, $H \times W = 192 \times 112$, Reynolds number $Re = 100$, and $N_s = 16$ uniformly distributed sensors over a sequence of 5000 consecutive snapshots.

The motivation is explicitly temporal. CNNs and UNets provide strong local spatial modeling but are typically frame-by-frame; RNNs and LSTMs suffer from vanishing or exploding gradients and memory degradation on long sequences; Transformers incur quadratic attention cost; and pure FNOs learn global spatial operators but do not by themselves resolve long-range temporal dependencies from sensor time series. FR-Mamba addresses this by combining a Fourier Neural Operator with a State Space Model instantiated as Mamba.

The architecture is dual-branch. The temporal branch is a Mamba-SSM that learns time-varying $B$, $C$, and $\Delta$ from embeddings and performs a linear-time scan after Zero-Order Hold discretization:

$$\bar{A} = e^{\Delta A}, \qquad \bar{B} = A^{-1}(e^{\Delta A} - I)B,$$

with discrete recurrence

$$h_k = \bar{A} h_{k-1} + \bar{B} x_k, \qquad y_k = C h_k.$$

The spatial branch uses FNO1d over the sensor dimension to capture non-local spatial correlations, and the decoded field is refined by a residual FNO2d:

$\hat{U} = \tilde{V} + \mathrm{FNO2d}(\tilde{V}).$

Fusion is gated through

$H \times W = 192 \times 112$0

followed by residual accumulation across stacked blocks.

Theoretical positioning is equally explicit. The SSM branch is presented as a learned global convolution with kernel $H \times W = 192 \times 112$1 and $H \times W = 192 \times 112$2 time and memory complexity in the sequence length $H \times W = 192 \times 112$3, while FNO layers provide global receptive fields through spectral convolutions with low-mode truncation. FR-Mamba therefore does not solve the Navier–Stokes equations directly; it learns a spatiotemporal operator from sparse vorticity measurements to dense fields, using operator-learning inductive biases and selective temporal memory (Long et al., 21 May 2025).

Training minimizes mean absolute error over space and time, without reported physics-based regularizers or spectral losses. The paper does not specify optimizer, learning rate schedule, batch size, epochs, normalization, or missing-sensor handling. The best ablation setting appears at $H \times W = 192 \times 112$4 and $H \times W = 192 \times 112$5, with average MAE $H \times W = 192 \times 112$6 and Max-AE $H \times W = 192 \times 112$7.

The empirical comparison is strong within the reported benchmark. Averaged across test intervals, FR-Mamba reports MAE $H \times W = 192 \times 112$8 of $H \times W = 192 \times 112$9 versus $Re = 100$0 for PFR-Transformer, $Re = 100$1 for LSTM-POD, and $Re = 100$2 for Voronoi UNet; average Max-AE $Re = 100$3 is $Re = 100$4 versus $Re = 100$5 for PFR-Transformer and $Re = 100$6 for Voronoi UNet. Ablations further show that removing both FNO1d and FNO2d degrades performance to MAE $Re = 100$7 and Max-AE $Re = 100$8, while FNO1d alone already yields large gains. The paper’s stated limitations concern sensor placement sensitivity, the absence of explicit noise modeling or uncertainty quantification, and untested generalization to unseen Reynolds numbers or geometries (Long et al., 21 May 2025).

3. FaRMamba for medical image segmentation

In "FaRMamba: Frequency-based learning and Reconstruction aided Mamba for Medical Segmentation" (Rong et al., 26 Jul 2025), FaRMamba is a frequency-aware, reconstruction-aided Vision Mamba architecture for medical segmentation. Its problem statement is framed by three persistent difficulties in clinical imaging: blurred lesion boundaries (LBA), loss of high-frequency details (LHD), and difficulty in modeling long-range anatomical structures (DC-LRSS). The paper argues that Vision Mamba already mitigates DC-LRSS through linear-complexity global dependency modeling, but that patch tokenization and one-dimensional recurrence introduce Local High-frequency Information Capture Deficiency (LHICD) and two-dimensional Spatial Structure Degradation (2D-SSD).

The architecture has four components: a shared Mamba encoder based on SS2D selective scanning, a Multi-Scale Frequency Transform Module (MSFM), a Self-Supervised Reconstruction Auxiliary Encoder (SSRAE), and a UMamba-style decoder. The encoder begins with a STEM layer, then patch embedding into $Re = 100$9 patches to produce $N_s = 16$0, followed by patch merging stages that reach $N_s = 16$1. SS2D extends image patches in four directions, processes them with selective scanning SSMs, and recombines them into 2D feature maps.

MSFM is inserted in shallow layers and explicitly restores high-frequency information before selective scanning. Within a unified framework, the paper evaluates DWT, FFT, and DCT variants. The DWT version uses first-order Daubechies wavelets, separates LL, LH, HL, and HH sub-bands, applies CBAM to each sub-band, and fuses high-frequency components before recombination with LL. The FFT version uses concentric circular masks with decreasing radii to isolate bands. The DCT version uses upper-triangular masks of increasing size. The stated function of all three is to isolate multi-band spectra, reconstruct spatial-domain feature maps, and reinject them so that high-frequency edges and textures are restored rather than attenuated (Rong et al., 26 Jul 2025).

SSRAE shares the same Mamba backbone as the segmentation branch and imposes a pixel-level reconstruction task. Its Degrade Block downsamples to $N_s = 16$2, applies a $N_s = 16$3 depthwise blur, and adds Gaussian noise with $N_s = 16$4. Reconstruction targets are high-frequency feature maps from the main encoder. Region attention masks are derived from segmentation labels, resized to feature resolution, flattened, and added as large negative values outside regions of interest during multi-head self-attention. The reconstruction objective combines L1, cosine similarity, and gradient losses, while the segmentation objective is a weighted sum of Dice loss and Cross-Entropy. The total loss uses a dynamic schedule in which reconstruction is progressively introduced and smoothed with an Exponential Moving Average.

The reported experiments cover CAMUS echocardiography, Mouse-cochlea MRI, and Kvasir-Seg endoscopy. On Mouse-cochlea, FaRMamba-FFT achieves Dice $N_s = 16$5 and MIoU $N_s = 16$6, improving on UMamba’s Dice $N_s = 16$7 and MIoU $N_s = 16$8. On CAMUS 2CH, FaRMamba-DWT reaches Dice $N_s = 16$9 and MIoU $B$0, exceeding UMamba’s Dice $B$1 and MIoU $B$2. On CAMUS 4CH, FaRMamba-DWT reports Dice $B$3 and MIoU $B$4, while nnUNet is cited at Dice $B$5 and MIoU $B$6. On Kvasir-Seg, FaRMamba-DCT attains Dice $B$7 and MIoU $B$8, compared with UMamba’s Dice $B$9 and MIoU $C$0 (Rong et al., 26 Jul 2025).

Ablations on CAMUS 2CH are used to separate the roles of MSFM and SSRAE. The base segmentation encoder gives Dice $C$1 and MIoU $C$2. Adding MSFM alone improves Dice to $C$3 for DWT, $C$4 for FFT, and $C$5 for DCT. Adding SSRAE without MSFM produces Dice around $C$6-$C$7. The full model yields the highest results: Dice $C$8 for DWT, $C$9 for FFT, and $\Delta$0 for DCT. The paper’s limitations are also explicit: transform choice is modality-dependent, SSRAE’s region attention relies on segmentation labels, and the dynamic reconstruction schedule is not fully explored across datasets (Rong et al., 26 Jul 2025).

4. Other papers linked to the label

The traffic restoration paper "Traffic Image Restoration under Adverse Weather via Frequency-Aware Mamba" states that the correct model name is Frequency-Aware Mamba, abbreviated FA-Mamba, and that the query term “FaRMamba” does not appear in the paper. The authors describe it as “most likely a typo or informal variant” for a framework built from a Dual-Branch Feature Extraction Block, a Prior-Guided Block, and an Adaptive Frequency Scanning Mechanism over wavelet sub-bands. The system targets rain, fog or haze, snow, and low-light degradation in traffic imagery, uses DWT decomposition into LL, LH, HL, and HH bands, and reports improvements over MambaIR on Snow100K subsets as well as a Raindrop PSNR/SSIM of $\Delta$1 (Pan et al., 3 Dec 2025).

The multivariate time-series forecasting paper "FMamba: Mamba based on Fast-attention for Multivariate Time-series Forecasting" states that “FaRMamba” colloquially captures the method’s core idea, but that the official name is FMamba. Technically, the model combines fast attention across variables with Mamba-based temporal scanning. It reports average MSE reductions versus S-Mamba and iTransformer across eight public datasets, including $\Delta$2 on PEMS03 and $\Delta$3 on PEMS07 relative to S-Mamba and iTransformer respectively, while maintaining low computational overhead (Ma et al., 2024).

The remote sensing super-resolution paper "Frequency-Assisted Mamba for Remote Sensing Image Super-Resolution" states that FaRMamba refers to the Frequency-Assisted Mamba framework, formally named FMSR. FMSR combines a Vision State Space Module, a Frequency Selection Module based on FFT-domain filtering, and a Hybrid Gate Module for local inductive bias. On AID, DOTA, and DIOR, it reports average PSNR/SSIM/LPIPS of $\Delta$4, compared with HAT-L’s $\Delta$5, while using about $\Delta$6 of HAT-L’s memory consumption and $\Delta$7 of its complexity (Xiao et al., 2024).

Taken together, these papers show that “FaRMamba” can denote an official model name, an alternative reference, or a retrospective shorthand. The term is therefore semantically unstable across domains.

5. Shared architectural patterns

Despite the naming ambiguity, several recurrent design motifs appear across the cited models. A first commonality is the use of Mamba or VSSM as the efficient long-range modeling core. FR-Mamba uses a selective SSM for temporal sequence modeling with $\Delta$8 time and memory complexity; medical FaRMamba uses SS2D selective scanning to model global anatomical structure; FA-Mamba adapts Mamba scanning paths to wavelet sub-bands; FMamba uses Mamba for temporal dependencies in forecasting; and FMSR uses VSSM for linear-complexity spatial aggregation (Long et al., 21 May 2025, Rong et al., 26 Jul 2025, Pan et al., 3 Dec 2025, Ma et al., 2024, Xiao et al., 2024).

A second recurring motif is the insertion of non-Mamba modules to compensate for perceived blind spots. In FR-Mamba, FNO1d and FNO2d provide global spatial operator learning absent from temporal SSMs. In medical FaRMamba, MSFM counters LHICD and SSRAE counters 2D-SSD. In FA-Mamba, wavelet-domain priors and AFSM recover texture distributions that “solely” spatial scanning does not capture. In FMamba, fast attention supplies explicit cross-variable mixing that channel-independent Mamba lacks. In FMSR, FFT-based frequency selection and pixel-wise hybrid gating supplement VSSM’s long-range scan (Long et al., 21 May 2025, Rong et al., 26 Jul 2025, Pan et al., 3 Dec 2025, Ma et al., 2024, Xiao et al., 2024).

A plausible synthesis is that the “FaRMamba” label tends to gather architectures that treat Mamba as a scalable backbone rather than as a complete solution. The auxiliary mechanisms differ by domain—operator learning for PDE-governed fields, frequency decomposition for medical and restoration tasks, fast attention for multivariate forecasting—but the structural logic is similar: preserve Mamba’s linear-time global modeling while reintroducing information pathways that would otherwise be suppressed or omitted.

6. Limitations, misconceptions, and future directions

The principal misconception is that FaRMamba identifies one established architecture. The cited papers do not support that interpretation. Instead, they document multiple domain-specific systems, some formally named FaRMamba, some only “also referred to as FaRMamba,” and some explicitly not bearing the term at all (Long et al., 21 May 2025, Rong et al., 26 Jul 2025, Pan et al., 3 Dec 2025, Ma et al., 2024, Xiao et al., 2024).

The technical limitations are similarly domain-specific. FR-Mamba does not report sensor-noise models, uncertainty quantification, or sensitivity analyses for uneven or clustered sensor placement, and it leaves unseen Reynolds numbers and geometries for future study (Long et al., 21 May 2025). Medical FaRMamba requires modality-sensitive transform selection and uses label-derived region attention in SSRAE, which the paper notes may introduce supervision bias and complicate transfer to unlabeled domains (Rong et al., 26 Jul 2025). FA-Mamba does not specify the wavelet basis, explicit AFSM mathematical criteria, or several block-level hyperparameters (Pan et al., 3 Dec 2025). FMamba notes that kernel choice and feature dimension $\Delta$9 in fast attention matter, and that its cross-variable interactions are computed independently at each time step (Ma et al., 2024). FMSR notes domain-shift sensitivity, training confined to $$\bar{A} = e^{\Delta A}, \qquad \bar{B} = A^{-1}(e^{\Delta A} - I)B,$$0 super-resolution with L1 supervision, and practical costs for ultra-high-resolution inference despite linear asymptotics (Xiao et al., 2024).

The forward-looking directions are correspondingly diverse but convergent in emphasis. FR-Mamba points to multi-physics coupling, 3D turbulent fields, adaptive sensor strategies, physics-informed constraints, and uncertainty quantification (Long et al., 21 May 2025). Medical FaRMamba proposes broader evaluation beyond medical segmentation and possible joint transform learning (Rong et al., 26 Jul 2025). FA-Mamba invites formalization of path selection and broader adverse-weather coverage (Pan et al., 3 Dec 2025). FMamba suggests adaptive kernels, dynamic variable graphs, and other structured cross-variable modules (Ma et al., 2024). FMSR highlights multi-band inputs, multi-task remote sensing, and temporal extensions such as video super-resolution (Xiao et al., 2024).

Within that broader landscape, FaRMamba is best understood as a name cluster around Mamba-based systems that augment linear-time state-space modeling with domain-specific corrective modules—frequency transforms, Fourier operators, auxiliary reconstruction, fast attention, or hybrid gating—rather than as a single, fixed architecture.