Patch Mamba: Patch-Token State-Space Modeling

Updated 8 July 2026

Patch Mamba is a patch-token formulation that decomposes visual, volumetric, temporal, or geometric inputs into ordered sequences for selective state-space modeling.
It leverages varied patch definitions, adaptive token construction, and multi-directional scanning to preserve spatial and temporal context across domains.
The approach achieves efficient linear-time processing, reducing computational overhead compared to quadratic self-attention in applications like remote sensing and medical imaging.

Searching arXiv for the provided paper and closely related "Patch Mamba" works to ground the article in current literature. arXiv search query: ([2410.13532](/papers/2410.13532)) OR "Patch Mamba" OR "Vision Mamba" Patch Mamba is a patch-token formulation of Mamba-based state space modeling in which visual, volumetric, temporal, or geometric inputs are decomposed into patch-like units, serialized into one or more sequences, and processed by selective state-space recurrences rather than quadratic self-attention. Taken together, recent work suggests that the term does not denote a single canonical architecture, but a family of designs that differ in patch definition, serialization order, directional scanning, and fusion strategy across remote sensing, medical imaging, time-series forecasting, point clouds, diffusion models, and tracking (Ren et al., 2024, Jeon, 7 Dec 2025, Bahri et al., 6 Mar 2025, Wang et al., 2024, Liu et al., 2 Jul 2025).

1. Conceptual scope and defining characteristics

In the recent literature, Patch Mamba is organized around a recurring computational pattern: construct patch tokens, map them to one-dimensional sequences, apply a Mamba or vision-adapted selective state-space block, and project the resulting states back into the target representation. In remote sensing, each spatial cell of a CNN feature map can function as a patch token and the full $H \times W$ grid is treated as a sequence of length $HW$ (Ren et al., 2024). In time-series forecasting, the normalized sequence can be segmented into non-overlapping patches of length $P$ , flattened, linearly projected, and then modeled by a Mamba-enhanced encoder (Jeon, 7 Dec 2025). In point clouds, patches can be local neighborhoods around farthest-point-sampled centers, ordered spectrally before Mamba processing (Bahri et al., 6 Mar 2025). In latent diffusion, Mamba can operate on latent patches extracted from a $4 \times 28 \times 28$ latent map (Wang et al., 2024).

This body of work also suggests that Patch Mamba is not restricted to explicit square image patches. Several systems treat feature-map cells, volumetric blocks, temporal segments, or graph-defined neighborhoods as the effective patches. RemoteDet-Mamba states that each spatial position of the CNN feature map acts as a patch token (Ren et al., 2024). CD-Mamba treats each position in a 2D feature map as a “patch token” corresponding to one spatial location at the current scale (Xue et al., 5 Sep 2025). MedSegMamba is explicitly 3D patch-based, processing $96 \times 96 \times 96$ MRI patches with stride 16 (Cao et al., 2024). This suggests that Patch Mamba is best understood as a tokenization-and-serialization regime for Mamba rather than as a fixed patch geometry.

A second defining characteristic is linear-time sequence modeling. Multiple papers contrast Mamba-based patch processing with Transformer attention on the grounds that Mamba scales linearly with sequence length, whereas self-attention is quadratic. That contrast is explicit in multimodal detection, cloud detection, time-series forecasting, and diffusion settings (Ren et al., 2024, Xue et al., 5 Sep 2025, Jeon, 7 Dec 2025, Wang et al., 2024). The technical consequence is that Patch Mamba is often introduced where long sequences, high resolutions, or dense prediction make attention expensive.

2. Patch construction and token formation

Patch construction varies substantially across domains. In image and feature-map settings, some systems use implicit patches defined by convolutional receptive fields. RemoteDet-Mamba first applies LayerNorm and Linear projection to intermediate CNN features, then depthwise convolution and SiLU, after which each spatial position becomes a patch token of dimension $C$ (Ren et al., 2024). CD-Mamba similarly flattens a feature map of shape $(b,c,H,W)$ into $(b,l,c)$ with $l=H \times W$ , so that each token is a spatial location whose receptive field grows with depth (Xue et al., 5 Sep 2025).

Other models make the patch structure explicit. AdaMamba segments the normalized time series into non-overlapping temporal patches of length $P$ , flattens each patch to $HW$ 0, and applies a linear projection to $HW$ 1 before positional encoding (Jeon, 7 Dec 2025). DCMamba partitions a 2D image into a regular $HW$ 2 grid of patches for weak-strong mixing augmentation, and extends the same idea to 3D patches for volumetric data (Li et al., 19 Aug 2025). MedSegMamba extracts heavily overlapping $HW$ 3 MRI patches from $HW$ 4 scans and reconstructs full-volume predictions by aggregating overlapping patch probabilities through a voting mechanism (Cao et al., 2024).

Several papers modify the patch itself rather than only its ordering. UAVD-Mamba introduces deformable tokens computed as

$HW$ 5

combining normal patches from standard convolution with adaptive patches from deformable convolution before Mamba processing (Li et al., 1 Jul 2025). UIS-Mamba goes further by learning offsets $HW$ 6 and scales $HW$ 7 for each patch, then defining deformed patches by bilinear interpolation before building a minimum spanning tree for scanning (Cong et al., 1 Aug 2025). These designs suggest that one branch of Patch Mamba research treats tokenization itself as a learnable geometric object.

In non-Euclidean settings, patch formation is graph-based. Spectral Informed Mamba builds patches around farthest-point-sampled centers using k-nearest neighbors, constructs a patch-connectivity graph, and then orders those patch tokens using low-frequency eigenvectors of the random-walk Laplacian (Bahri et al., 6 Mar 2025). This is a direct adaptation of Patch Mamba to domains without a canonical raster order.

3. Serialization, scan order, and directional processing

Serialization is the most contested design choice in Patch Mamba. Since Mamba is a sequential model, the mapping from a 2D, 3D, or graph-structured patch set to a 1D order materially affects which dependencies are easy or hard to model. A systematic MRI segmentation benchmark over 21 scan strategies showed that scan order is a statistically significant factor, with Friedman test $HW$ 8 and $HW$ 9, and with performance varying by as much as 27 Dice points (Hardan et al., 15 Jul 2025). The same study found that spatially contiguous horizontal and vertical rasters consistently outperform disjointed diagonal scans (Hardan et al., 15 Jul 2025). A separate visualization study of vision-based Mamba models similarly reported that patches close in the input sequence generally exhibit similar attention patterns and that changing the patch order reshapes those patterns while preserving comparable ImageNet accuracy (Wang et al., 28 Feb 2025).

Many Patch Mamba designs therefore use multi-directional scans. RemoteDet-Mamba performs quad-directional selective scanning by flattening 2D feature maps in four directions, processing each with an S6 block, unflattening, and summing the directional outputs (Ren et al., 2024). VMamba-derived systems in medical segmentation and multimodal detection also rely on multiple scan routes to approximate 2D context (Li et al., 19 Aug 2025, Li et al., 1 Jul 2025). Patch-MoE Mamba identifies a limitation of pixel-wise directional scanning—namely that it disrupts local 2D structure—and replaces it with hierarchical patch-ordered scanning, where pixels within a patch remain contiguous in the sequence (Adame et al., 18 May 2026).

A parallel line of work argues that raster-like serialization is itself the main bottleneck. GlobalMamba explicitly critiques patch-based flattening, noting that flattened tokens contain spatially confined information and do not align well with causal sequence modeling; it replaces patch serialization with frequency-based global image serialization using DCT, band-wise inverse transforms, and low-to-high frequency causal ordering (Wang et al., 2024). FractalMamba++ replaces raster ordering with Hilbert-curve serialization, formalizes spatial distortion through the Structure Distortion Score

$P$ 0

and couples fractal traversal with Cross-State Routing and Positional-Relation Capture to improve locality preservation and resolution scaling (Li et al., 20 May 2025).

Recent work also attempts to recover bidirectional or non-causal interactions without expensive multi-scan rearrangement. SF-Mamba introduces auxiliary patch swapping to encode bidirectional information flow under a unidirectional scan and batch folding with periodic state reset for improved GPU parallelism, explicitly framing prior multi-scan strategies as inefficient under short token lengths typical of vision (Yoshimura et al., 17 Mar 2026). This suggests that Patch Mamba research has shifted from merely finding any workable serialization to optimizing the inductive bias and systems efficiency of the scan itself.

4. Hybridization, fusion, and expert routing

Patch Mamba is rarely deployed as a pure SSM stack. Instead, recent systems use Mamba within broader hybrid pipelines that combine convolution, attention, expert routing, or task-specific conditioning. RemoteDet-Mamba places Mamba only inside a Cross-modal Fusion Mamba block, keeping a Siamese CSPDarkNet53 backbone and a YOLO-style detector while performing patch-level RGB–TIR fusion through sequence addition and quad-directional scanning (Ren et al., 2024). UAVD-Mamba likewise couples deformable token Mamba blocks, a Fusion Mamba Block, cross-enhanced spatial attention, cross-channel attention, and a YOLOv11-style neck (Li et al., 1 Jul 2025).

Mixture-of-experts mechanisms appear in several Patch Mamba variants. AdaMamba uses patch-wise embeddings followed by a Mamba-enhanced Transformer layer with a mixture-of-experts feed-forward module, routing patch tokens through softmax-gated experts to handle heterogeneous temporal dynamics (Jeon, 7 Dec 2025). Patch-MoE Mamba introduces an MoE-based directional fusion module that adaptively combines four directional Mamba outputs, a learnable concatenation expert, and residual directional aggregation rather than simply summing all scans (Adame et al., 18 May 2026). In both cases, the expert mechanism is positioned after or around patch-sequence modeling, not as a replacement for it.

Conditioning strategies are similarly diverse. DiffMa conditions latent patch sequences on CT-derived embeddings and soft masks, integrates Cross-Sequence Attention into Mamba, and uses spiral selective scanning to preserve spatial continuity of latent patches in CT-to-MRI conversion (Wang et al., 2024). TrackingMiM uses a Mamba-in-Mamba encoder in which spatial Mamba processes per-frame patch sequences and temporal Mamba processes the corresponding patch positions across frames; the template frame is encoded as a query token and reused in every scan (Liu et al., 2 Jul 2025). UIS-Mamba modulates hidden-state propagation by an Ncut-derived foreground/background partition, weakening background patches through a suppression factor $P$ 1 and thereby focusing state propagation on instance regions (Cong et al., 1 Aug 2025).

These systems indicate that Patch Mamba is often most effective when the state-space core is treated as a sequence-processing substrate inside a more structured architecture. A plausible implication is that patch ordering alone is seldom sufficient; token construction, conditioning, and fusion determine whether the linear-time recurrence captures the right dependencies for the target task.

5. Representative application domains

Patch Mamba has been instantiated across multiple data regimes, with domain-specific meanings of “patch.”

Domain	Patch definition	Representative mechanism
Remote sensing detection	CNN feature-map cells as patch tokens	Quad-directional selective scanning fusion (Ren et al., 2024)
Time-series forecasting	Non-overlapping temporal segments of length $P$ 2	Patch embedding + MambaMoELayer (Jeon, 7 Dec 2025)
Medical image segmentation	Regular image patches, $P$ 3 volumetric patches, or patch-ordered scans	Patch-ordered scanning, SS3D, weak-strong patch mixing (Adame et al., 18 May 2026, Cao et al., 2024, Li et al., 19 Aug 2025)
Point clouds	FPS+kNN local neighborhoods	Spectral ordering via Laplacian eigenvectors (Bahri et al., 6 Mar 2025)
Diffusion and tracking	Latent patches or frame patches	Spiral scan, nested spatial/temporal Mamba (Wang et al., 2024, Liu et al., 2 Jul 2025)

In remote sensing, Patch Mamba is used to enhance small-object and dense-object reasoning. RemoteDet-Mamba reports that CFM sits between the two modality feature extraction networks and achieves deep patch-level global feature fusion between RGB and infrared modalities (Ren et al., 2024). CD-Mamba targets cloud detection by combining shallow convolution with deeper Cloud-SMB blocks that model long-term patchwise dependencies through four SSM scan directions, achieving $P$ 4 mIoU on Biome and $P$ 5 mIoU on GF1-WHU while remaining extremely lightweight (Xue et al., 5 Sep 2025).

In medical imaging, the term often carries stronger geometric or volumetric content. MedSegMamba is explicitly a 3D patch-based hybrid CNN-Mamba model operating on $P$ 6 patches, reaching overall DSC $P$ 7, VS $P$ 8, and ASSD $P$ 9 for subcortical brain segmentation (Cao et al., 2024). DCMamba combines patch-level weak-strong mixing augmentation, scan-direction diversity, and uncertainty-weighted contrastive learning, improving over Semi-Mamba-UNet by $4 \times 28 \times 28$ 0 Dice on Synapse 2D with 20% labels (Li et al., 19 Aug 2025). Patch-MoE Mamba pushes this line further by replacing pixel-wise scanning with hierarchical patch-ordered scanning and adaptive directional fusion, reaching $4 \times 28 \times 28$ 1 Dice on ETIS and $4 \times 28 \times 28$ 2 Dice on ISIC 2018 (Adame et al., 18 May 2026).

In time-series modeling, Patch Mamba denotes temporal rather than spatial patching. AdaMamba normalizes the input, segments it into non-overlapping temporal patches, and models those patch tokens with a Mamba-enhanced Transformer and MoE feed-forward stack; it achieves MSE $4 \times 28 \times 28$ 3 on ETTh1, $4 \times 28 \times 28$ 4 on ETTh2, $4 \times 28 \times 28$ 5 on ETTm1, and $4 \times 28 \times 28$ 6 on ETTm2, with Weather performance essentially tied with iTransformer at $4 \times 28 \times 28$ 7 versus $4 \times 28 \times 28$ 8 (Jeon, 7 Dec 2025).

For non-grid data, Spectral Informed Mamba shows that Patch Mamba can be defined intrinsically. It constructs patch graphs on point clouds, canonicalizes Laplacian eigenvectors, and derives SAST and HLT traversals for classification and segmentation, improving ScanObjectNN OBJ-BG accuracy from $4 \times 28 \times 28$ 9 to $96 \times 96 \times 96$ 0 from scratch and from $96 \times 96 \times 96$ 1 to $96 \times 96 \times 96$ 2 with pretraining (Bahri et al., 6 Mar 2025). This is notable because it demonstrates that patch tokens need not come from rectangular decompositions.

6. Limitations, debates, and research trajectory

A recurring limitation is that patch serialization remains a lossy 2D-to-1D or higher-dimensional-to-1D reduction. GlobalMamba explicitly argues that conventional patch flattening ignores intrinsic 2D structural correlations and produces tokens that lack comprehensive global context (Wang et al., 2024). FractalMamba++ reaches a similar conclusion from a locality-preservation perspective and adds that high-resolution scaling fails when long-range dependency propagation weakens along long recurrent chains (Li et al., 20 May 2025). Flatten Wisely shows that even within a single backbone, the wrong scan order can be catastrophically bad (Hardan et al., 15 Jul 2025). These results collectively challenge the assumption that patch order is a mere implementation detail.

Another limitation is that linear complexity does not automatically imply high throughput. SF-Mamba argues that Mamba can be relatively slow under short token lengths common in vision, and that previous multi-scan strategies suffer from suboptimal scan designs and frequent data rearrangement (Yoshimura et al., 17 Mar 2026). CD-Mamba and RemoteDet-Mamba both emphasize speed advantages from Mamba’s serial processing, but the SF-Mamba critique indicates that realized efficiency depends heavily on kernel implementation, scan topology, and batching strategy rather than on asymptotic complexity alone (Ren et al., 2024, Xue et al., 5 Sep 2025, Yoshimura et al., 17 Mar 2026).

A further debate concerns whether patchification should remain fixed. UIS-Mamba and UAVD-Mamba argue that fixed patch boundaries can be harmful in segmentation or small-object detection, motivating dynamic patch deformation and deformable token generation (Cong et al., 1 Aug 2025, Li et al., 1 Jul 2025). Spectral Informed Mamba makes a related argument for point clouds, where naive grid traversals are view-dependent and fail to respect surface geometry (Bahri et al., 6 Mar 2025). This suggests a broader shift from static patch tokens toward geometry-aware or data-adaptive tokenization.

Finally, several papers imply that the future of Patch Mamba lies in stronger structure rather than in larger raw SSM stacks. Hierarchical patch-ordered scanning, MoE fusion, fractal traversal, spectral traversal, spiral or tree scans, retrieval-augmented conditioning, and adaptive normalization all modify the way patches are formed, ordered, or reused rather than merely increasing depth (Adame et al., 18 May 2026, Li et al., 20 May 2025, Bahri et al., 6 Mar 2025, Wang et al., 2024, Liu et al., 2 Jul 2025). A plausible implication is that Patch Mamba will remain a design space rather than converging quickly to a single dominant template. What unifies that space is the idea that Mamba is most effective when the patch sequence itself is engineered to preserve the task’s underlying topology.