RadioMamba: Domain-Specific Selective SSM Applications

• RadioMamba is a family of domain-adapted selective state-space models that capture long-range temporal or spatial dependencies efficiently.
• It encompasses applications in MRI-to-CT synthesis, wireless communications, radio map construction, and radar-based human activity recognition.
• Hybrid designs integrate SSMs with convolutional and attention mechanisms to improve accuracy and reduce inference time across domains.

Searching arXiv for "RadioMamba" and closely related papers to ground the article in current literature. RadioMamba denotes several domain-specific adaptations of Mamba-style selective state-space models rather than a single standardized architecture. In current arXiv usage, the name identifies a Mamba-driven pipeline for MRI-to-CT synthesis in MRI-only radiotherapy planning, a label for applying selective state-space models to wireless communications and networking, and a hybrid Mamba-UNet for radio map construction; a closely related but separately named line, RadMamba, targets radar-based human activity recognition (Barmpounakis et al., 24 Mar 2026, Zhang et al., 1 Aug 2025, Jia et al., 28 Jul 2025, Wu et al., 16 Apr 2025). Across these usages, the shared premise is that selective SSMs can model long-range temporal or spatial dependencies with linear-time or near-linear-time execution, while each instantiation introduces domain-specific tokenization, architectural hybridization, training objectives, and evaluation protocols.

1. Nomenclature and scope

The term has been used in multiple, technically distinct contexts.

Name	Domain	Reported formulation
RadioMamba	MRI-only radiotherapy planning	Mamba-driven pipeline for cross-modality MRI-to-CT synthesis adapting U-Mamba and SegMamba
RadioMamba	Wireless communications and networking	selective state-space architecture used in two frameworks: replacement of traditional algorithms, and enabler of novel paradigms
RadioMamba	Radio map construction	hybrid Mamba-UNet architecture with MambaConvBlock
RadMamba	Radar-based HAR	radar-based micro-Doppler-oriented Mamba SSM

This suggests that “RadioMamba” functions as a family label or convergent naming pattern rather than a canonical architecture with a single reference implementation. The commonality is methodological: each work places a selective SSM at the center of a system that must preserve long-range structure under stringent efficiency constraints, whether the structure is volumetric anatomy, channel dynamics, propagation geometry, or micro-Doppler evolution (Barmpounakis et al., 24 Mar 2026, Zhang et al., 1 Aug 2025, Jia et al., 28 Jul 2025, Wu et al., 16 Apr 2025).

2. Selective state-space foundations

The MRI synthesis formulation presents a discrete linear time-invariant SSM operating along one spatial dimension, such as slice or depth. Its basic form is

$$\mathbf{s}_t = \mathbf{A}\,\mathbf{s}_{t-1} + \mathbf{B}\,\mathbf{u}_t,$$

$$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$

Here, $\mathbf{u}_t$ is the input feature at slice index $t$, $\mathbf{s}_t$ is the latent state, and $\mathbf{h}_t$ is the SSM output, which is reshaped and fused back into the volumetric feature map. The implementation uses a diagonal plus low-rank parameterization of $\mathbf{A}$ for $O(N)$ time and memory, and interleaves SSM layers with 3D pointwise and depthwise convolutions to blend local and global context (Barmpounakis et al., 24 Mar 2026).

The wireless formulation begins from a continuous-time state-space system,

$$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$

then discretizes it with a learnable step $\Delta_t$ to obtain

$$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$0

The same work also gives a global-convolution view,

$$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$1

and emphasizes that a parallel scan can compute $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$2 in linear time, yielding $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$3 rather than $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$4 complexity in sequence length $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$5 (Zhang et al., 1 Aug 2025).

The radio-map construction variant adapts Mamba to 2D feature maps by raster-scanning the map into a sequence. Within its SS2D-Mamba branch, the selective dynamics are parameterized by

$$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$6

with recurrence

$$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$7

The branch is bidirectional: a forward pass is combined with a reversed-sequence pass and then reshaped back into a 2D tensor (Jia et al., 28 Jul 2025).

A recurring pattern across these formulations is input-dependent modulation of transition, input, or output projections. This suggests that Mamba is being used not merely as a replacement for recurrence, but as a mechanism for adaptive memory control under nonstationary spatial or temporal structure.

3. RadioMamba in MRI-only radiotherapy planning

In "Mamba-driven MRI-to-CT Synthesis for MRI-only Radiotherapy Planning" (Barmpounakis et al., 24 Mar 2026), RadioMamba is the designation for a pipeline that generates synthetic CT (sCT) volumes from input MRI by adapting U-Mamba and SegMamba, both originally proposed for segmentation. The stated goals are to eliminate repeated CT scans, reduce ionizing dose, and avoid inter-modality registration errors.

The 3D U-Mamba variant retains the U-Net encoder-decoder topology of nnU-Net but replaces every convolutional block with a Mamba block comprising a 3D pointwise convolution, a selective SSM along the slice or depth axis, and a gating mechanism such as FiLM-style re-weighting before residual addition. SegMamba is described as a hybrid pyramid alternating convolution and Mamba blocks at multiple scales, using 3×3×3 convolutions for local modeling, SSM layers for sequence modeling in depth, and skip-connections for encoder-decoder fusion. The paper attributes to these designs the ability to capture complex volumetric features and long-range dependencies while maintaining fast inference times (Barmpounakis et al., 24 Mar 2026).

Training used a curated subset of SynthRAD2025 with 461 training and 52 held-out test MRI–CT pairs across abdomen, head-and-neck, and thorax from three European centers. All volumes were rigidly registered with Elastix, resampled to $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$8 mm, cropped with a 10-pixel body margin, MRI z-score normalized per scan, and CT clipped to $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$9 HU and then z-score normalized globally. Models consumed 3D patches of size $\mathbf{u}_t$0 with at least 70% body coverage. The loss was staged: for epochs $\mathbf{u}_t$1, weighted MAE in HU emphasized bone and soft tissue while de-emphasizing air; for epochs $\mathbf{u}_t$2, SSIM loss and AFP loss using TotalSegmentator embeddings were added. Optimization used AdamW with initial learning rate $\mathbf{u}_t$3, polynomial decay, and 500 epochs, whereas the nnU-Net baseline was trained for 1000 epochs with SGD (Barmpounakis et al., 24 Mar 2026).

Evaluation combined HU-based image similarity with segmentation-based geometric consistency. The HU metrics were MAE, RMSE, PSNR, and MS-SSIM. Segmentation-based metrics, derived from TotalSegmentator, were Dice Similarity Coefficient and 95th percentile Hausdorff Distance. On the test set, SegMamba reported MAE $\mathbf{u}_t$4, PSNR $\mathbf{u}_t$5, MS-SSIM $\mathbf{u}_t$6, DSC $\mathbf{u}_t$7, and HD95 $\mathbf{u}_t$8; U-Mamba reported MAE $\mathbf{u}_t$9, PSNR $t$0, MS-SSIM $t$1, DSC $t$2, and HD95 $t$3. The nnU-Net baseline reported MAE $t$4, PSNR $t$5, MS-SSIM $t$6, DSC $t$7, and HD95 $t$8, while U-Net and SwinUNETR were worse on voxel-level metrics. Inference times on an NVIDIA A100 were reported as approximately $t$9–$\mathbf{s}_t$0 s per $\mathbf{s}_t$1 patch, with full-volume inference under 2 s (Barmpounakis et al., 24 Mar 2026).

The paper’s own interpretation is balanced. Benefits include strong long-range modeling, improved HU fidelity, lower capacity than nnU-Net, and fast inference suitable for online adaptive planning. Limitations include slightly lower PSNR, MS-SSIM, and DSC than nnU-Net, as well as new SSM hyperparameters such as sequence length that require careful tuning. Integration into MRI-only workflows is described as MRI acquisition $\mathbf{s}_t$2 z-score plus patching $\mathbf{s}_t$3 U-Mamba or SegMamba inference $\mathbf{s}_t$4 HU-clipping plus DICOM conversion $\mathbf{s}_t$5 dose calculation (Barmpounakis et al., 24 Mar 2026).

4. RadioMamba in wireless communications and networking

"Mamba for Wireless Communications and Networking: Principles and Opportunities" uses “RadioMamba” as a label for applying selective SSMs to radio and wireless systems (Zhang et al., 1 Aug 2025). The work frames Mamba as a model for jointly handling rich temporal structure, such as time-varying fading and user mobility, and complex spatial coupling, such as multi-antenna MIMO channels and ultra-dense node graphs.

Two application frameworks are proposed. The first is replacement of traditional signal-processing modules. In this setting, Mamba blocks replace iterative or hand-crafted algorithms in tasks such as channel estimation, beamforming, precoding, and message-passing detection. The paper gives channel estimation as an example in which compressed-sensing plus AMP loops are replaced by a Mamba SSM ingesting pilot sequences and outputting refined CSI, and beamforming or precoding as a case where interior-point or fixed-point updates are supplanted by Mamba tracking of inter-user interference patterns. The stated trade-off is that interpretability of closed-form solvers is lost, but inference speed improves by an order of magnitude in ultra-dense scenarios and MSE or throughput matches or exceeds traditional baselines under dynamic fading (Zhang et al., 1 Aug 2025).

The second framework is the enabler of novel paradigms. Examples include intelligent resource management, in which Mamba observes long-term traffic demand and inter-cell coupling to output power or spectrum allocation in one shot, and joint source/channel coding and decoding, in which selective SSM blocks are inserted into both semantic and channel components so that optimization can target semantic fidelity such as BLEU score rather than separate bit-error metrics (Zhang et al., 1 Aug 2025).

The intelligent resource-allocation case study considers a multi-user MISO downlink with $\mathbf{s}_t$6 single-antenna users and an $\mathbf{s}_t$7-antenna base station, optimizing

$\mathbf{s}_t$8

subject to $\mathbf{s}_t$9 and $\mathbf{h}_t$0. A GNN-Mamba architecture begins with graph attention over the user-base-station bipartite graph and replaces half of the attention layers by Mamba blocks. The reported complexity scaling is

$\mathbf{h}_t$1

with relative energy-efficiency loss $\mathbf{h}_t$2 for $\mathbf{h}_t$3 up to 50. The paper reports that as $\mathbf{h}_t$4 grows from 15 to 50, GNN inference time rises by $\mathbf{h}_t$5, whereas GNN-Mamba time remains essentially flat; the energy-efficiency ratio stays above 98% for $\mathbf{h}_t$6 and above 85% even at $\mathbf{h}_t$7 (Zhang et al., 1 Aug 2025).

The joint source-channel decoding case study modifies DeepSC by replacing an intermediate self-attention block on both transmitter and receiver sides with Mamba blocks. At low SNR, 0–3 dB, Transformer+Mamba lags slightly, with 49.9% versus 52.1% BLEU at 0 dB, but at moderate-to-high SNR, 9–18 dB, it outperforms DeepSC by 3–4 points, including 92.4% versus 90.6% at 18 dB, while reducing per-sequence latency by approximately 25% (Zhang et al., 1 Aug 2025).

The work closes with open problems that include joint task scheduling in MEC, interference dynamics in heterogeneous networks, V2X cooperative perception and fusion, and collaborative UAV swarm control. In this literature, RadioMamba is therefore less a single model than a general deployment strategy for selective SSMs in wireless stacks.

5. RadioMamba in radio map construction

"RadioMamba: Breaking the Accuracy-Efficiency Trade-off in Radio Map Construction via a Hybrid Mamba-UNet" defines RadioMamba as a U-Net–style encoder-decoder in which standard convolutional blocks are replaced by a MambaConvBlock (Jia et al., 28 Jul 2025). The task is radio map construction for real-time and accurate spatial channel information in 6G settings.

The network takes an input of size $\mathbf{h}_t$8 with three channels: a static obstacle map, an optional dynamic-obstacle map, and a one-hot transmitter location. The encoder has three resolution levels, $\mathbf{h}_t$9, each containing two MambaConvBlocks followed by a $\mathbf{A}$0 strided convolution that doubles channel count. The bottleneck at $\mathbf{A}$1 contains two more MambaConvBlocks. The decoder mirrors the encoder through three upsampling stages, each consisting of transposed $\mathbf{A}$2 convolution, concatenation with the corresponding skip feature map, fusion by a $\mathbf{A}$3 convolution, and refinement by two MambaConvBlocks. The output is a final $\mathbf{A}$4 convolution producing a single-channel radio-map prediction (Jia et al., 28 Jul 2025).

Each MambaConvBlock has two parallel branches. The local branch is a ResidualConvBranch with depthwise $\mathbf{A}$5 convolution, GELU, and pointwise $\mathbf{A}$6 convolution, followed by residual addition. The reported MAC cost drops from standard convolution cost to depthwise-plus-pointwise cost, which for kernel size $\mathbf{A}$7 and large $\mathbf{A}$8 is about $\mathbf{A}$9 of a standard convolution. The global branch is an SS2D-MambaBranch: LayerNorm, flattening to a sequence of length $O(N)$0, bidirectional Mamba, summation of forward and backward outputs, and unflattening. Final fusion is simply

$O(N)$1

The paper’s complexity analysis states that self-attention on $O(N)$2 tokens costs $O(N)$3 flops and $O(N)$4 memory, whereas Mamba’s selective SSM can be implemented with dominant cost $O(N)$5, and the total MambaConvBlock cost is $O(N)$6, linear in the number of pixels rather than quadratic as in a Transformer block (Jia et al., 28 Jul 2025).

Training used the RadioMapSeer benchmark with 56,000 samples. Each sample is a $O(N)$7 grid containing static building layout, transmitter one-hot map, and optional dynamic-obstacle map; outputs are path-loss in dB, min–max normalized to $O(N)$8. Dynamic maps are synthesized by randomly placing vehicle shapes along roads, and no further geometric augmentations are used. Optimization employed AdamW with learning rate $O(N)$9, weight decay $$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$0, cosine annealing over 200 epochs, mixed precision on NVIDIA A40 GPUs, and batch size 16. The loss was

$$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$1

where $$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$2 is an $$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$3 difference on Sobel edges (Jia et al., 28 Jul 2025).

On static radio maps, RadioMamba reported NMSE $$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$4, RMSE $$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$5, SSIM $$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$6, and PSNR $$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$7; on dynamic radio maps, it reported NMSE $$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$8, RMSE $$\frac{d h(t)}{dt} = A\,h(t) + B\,u(t), \qquad y(t) = C\,h(t),$$9, SSIM $\Delta_t$0, and PSNR $\Delta_t$1. The efficiency table reports $\Delta_t$2 s per map, $\Delta_t$3 M parameters, and $\Delta_t$4 MB GPU memory, alongside an “improvement over RadioDiff” of $\Delta_t$5 inference time, $\Delta_t$6 parameters, and $\Delta_t$7 GPU memory. The abstract summarizes this as higher accuracy than existing methods, including diffusion models, while operating nearly 20 times faster and using only 2.9% of the model parameters (Jia et al., 28 Jul 2025).

The paper attributes these gains to a hybridization principle: the Mamba branch captures global propagation dependencies such as reflections and diffractions, while the convolutional branch preserves local detail such as sharp shadow edges. Applications explicitly named include network digital twin updates, UAV trajectory planning, and dynamic spectrum management and interference coordination in dense 6G cells (Jia et al., 28 Jul 2025).

6. RadMamba as a related radar-oriented line

A separate but adjacent literature uses the name RadMamba rather than RadioMamba. "RadMamba: Efficient Human Activity Recognition through Radar-based Micro-Doppler-Oriented Mamba State-Space Model" targets radar-based HAR from micro-Doppler spectrograms (Wu et al., 16 Apr 2025).

The signal representation begins with continuous-wave or FMCW returns and forms a micro-Doppler spectrogram through the STFT,

$\Delta_t$8

yielding a real spectrogram tensor $\Delta_t$9. The paper reports empirical sparsity in CI4R of approximately 87% zero-energy pixels. A small Conv2D $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$00 BatchNorm $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$01 MaxPool2D $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$02 MaxPool1D block, called Chan-DS, first fuses channels and reduces spatial dimensions (Wu et al., 16 Apr 2025).

RadMamba then replaces standard multi-head self-attention in Vision Mamba with bidirectional SSM branches but adds three radar-specific choices: Doppler-Aligned Segmentation, which extracts patches of size $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$03 along the full Doppler axis and one time-bin width; Conv1D projections in the rMamba blocks; and a residual block structure with LayerNorm, forward and backward SSM branches, SiLU gating, and a final Conv1D projection. The discrete-time SSM is written as

$$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$04

with input-dependent parameterization of $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$05, $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$06, and $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$07 through learned projections and Softplus (Wu et al., 16 Apr 2025).

The reported best configurations are dataset-specific. For DIAT, the best RadMamba has dimension 64, state dimension $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$08, $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$09, 21.7 k parameters, and 145.6 M FLOP per inference. For CI4R, it uses dimension 96, $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$10, $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$11, 71.4 k parameters, and 8.8 M FLOP. For UoG2020, it uses dimension 24, $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$12, $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$13, 6.7 k parameters, and 1.0 M FLOP (Wu et al., 16 Apr 2025).

Performance is reported over 10 random seeds. On DIAT, RadMamba attains $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$14 accuracy with 21.7 k parameters and 145.6 M FLOP, matching the 99.8% of ActivityMamba but with only 1/400 of its parameters. On CI4R, it attains $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$15, versus 92.0%–92.1% for ViT-based baselines, while being about 1/10 of their size and 20× lighter in FLOP. On UoG2020, it attains $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$16 with only 6.7 k parameters and 1.0 M FLOP, surpassing larger comparison models by at least 3%. The paper further reports convergence in approximately 20–30 epochs and ablations showing that the combination of Conv1D projections, Doppler-aligned patching, and adaptive downsampling is required for top performance (Wu et al., 16 Apr 2025).

Although not titled RadioMamba, RadMamba demonstrates the same design tendency visible in the other lines: domain-aware restructuring of sequence layout and a preference for lightweight SSM-based hybrids over quadratic attention.

7. Cross-domain interpretation and open directions

Across the cited works, several design regularities recur. First, long-range dependency modeling is treated as indispensable, but almost never in isolation. The MRI pipeline interleaves SSM layers with 3D pointwise and depthwise convolutions; the radio-map model uses a parallel convolutional branch and a parallel Mamba branch inside every MambaConvBlock; the wireless case study combines graph attention with Mamba; and RadMamba supplements bidirectional SSMs with Conv1D projections and Doppler-aligned segmentation (Barmpounakis et al., 24 Mar 2026, Zhang et al., 1 Aug 2025, Jia et al., 28 Jul 2025, Wu et al., 16 Apr 2025). This suggests that successful RadioMamba-style systems are typically hybrid rather than purely state-space.

Second, efficiency claims are central but domain-specific. In MRI synthesis, inference is under 2 s per full volume and parameter count is smaller than nnU-Net; in radio map construction, inference is $$\mathbf{h}_t = \mathbf{C}\,\mathbf{s}_t + \mathbf{D}\,\mathbf{u}_t.$$17 s per map with 8.60 M parameters and reported 95% time reduction versus RadioDiff; in wireless case studies, Mamba changes scaling behavior from quadratic to linear or empirically flat in user count, or reduces per-sequence latency by about 25%; in radar HAR, competitive or superior accuracy is achieved with parameter counts as low as 6.7 k (Barmpounakis et al., 24 Mar 2026, Zhang et al., 1 Aug 2025, Jia et al., 28 Jul 2025, Wu et al., 16 Apr 2025).

Third, the limitations are likewise structured by application. The MRI work reports slightly lower PSNR, MS-SSIM, and DSC than nnU-Net and notes additional SSM hyperparameters. The wireless overview acknowledges loss of the interpretability associated with closed-form solvers in module-replacement settings. The radio-map paper identifies future extensions including 3D volumetric maps, pruning and quantization, federated training, and space-filling-curve flattening to reduce 2D anisotropy. The MRI paper calls for institution-specific fine-tuning, end-to-end QA with dose recalculation, and prospective clinical validation, while the wireless paper lists MEC scheduling, heterogeneous-network interference prediction, V2X cooperative fusion, and collaborative UAV swarm control as open directions (Barmpounakis et al., 24 Mar 2026, Zhang et al., 1 Aug 2025, Jia et al., 28 Jul 2025).

Taken together, the current literature indicates that RadioMamba is best understood as a domain-adapted Mamba design paradigm. Its core proposition is stable across fields—selective SSMs for long-range dependency modeling with efficient execution—but its concrete meaning depends on the application domain, the geometry of the data, and the operational metric that matters most, such as HU fidelity, BLEU score, energy efficiency, path-loss reconstruction quality, or radar HAR accuracy.