Serialized Point Mamba Architectures

Updated 21 February 2026

Serialized Point Mamba is a framework that serializes unordered 3D point clouds into sequences while preserving spatial locality for effective state space modeling.
It employs diverse spatially-aware techniques such as space-filling curves, spectral ordering, and learnable permutations to optimize tasks like classification, segmentation, registration, and generation.
This approach achieves linear computational complexity and state-of-the-art accuracy across benchmarks by integrating hybrid pipelines and adaptive serialization strategies.

Serialized Point Mamba refers to the class of architectures and methodologies that enable Mamba state space models (SSMs) to operate effectively on point clouds and irregular 3D geometric data by serializing unordered sets or spatial arrays into sequences. Central to these approaches is the use of spatially-aware serialization strategies—such as space-filling curves, spectral orders, grid/voxel discretization, and learnable permutations—that preserve geometrical locality and enable linear-complexity sequence modeling with SSMs. Serialized Point Mamba delivers compelling accuracy and efficiency on classification, segmentation, registration, and generative modeling tasks across real-world point cloud benchmarks (Liu et al., 2024, Bahri et al., 6 Mar 2025, Wang et al., 2024, Lin et al., 23 Jul 2025, Liu et al., 16 Jun 2025, Liu et al., 17 Mar 2025, Zhang et al., 2024, Li et al., 20 May 2025, Zha et al., 27 May 2025, Zhang et al., 7 Jun 2025).

1. Principles of Mamba and the Need for Serialization

Mamba is a structured state space model designed for sequence modeling with linear computational and memory complexity in input length. Unlike self-attention (Transformers), SSMs require input data to be fed as a sequence with strict causality or scan order. However, point clouds and spatial arrays are unordered sets or irregular grids with no canonical sequence, requiring serialization that both defines a sequence order and preserves spatial relationships critical for geometric reasoning.

Key desiderata for point cloud serialization include:

Total ordering: Every point is assigned a unique position in the sequence.
Spatial locality: Points close in 3D space remain close in the 1D sequence.
Task/task-agnostic flexibility: Orderings may be adapted for classification, segmentation, registration, or generation, potentially through learned permutation or task-aware policies.
Computational tractability: The serialization and SSM operations must maintain linear time/space scaling in the number of points or patches (Liu et al., 2024, Bahri et al., 6 Mar 2025, Zha et al., 27 May 2025).

2. Spatial Serialization Strategies

Multiple serialization strategies have been designed and evaluated for Serialized Point Mamba. These include:

(a) Space-Filling Curves

Space-filling curves map multidimensional coordinates to a one-dimensional sequence while attempting to preserve locality.

Morton/Z-order (Octree ordering): Interleaves the bits of discretized x, y, z coordinates, producing a sequence in which spatial neighbors differ only in low-order bits. Used in (Liu et al., 2024, Liu et al., 16 Jun 2025, Liu et al., 17 Mar 2025).

$K(x, y, z) = \sum_{i=0}^{d-1}(2^{3i+2}x_i + 2^{3i+1}y_i + 2^{3i}z_i)$

Hilbert/Trans-Hilbert: Recursive fractal curves achieving higher locality preservation than Z-order, especially at high resolutions and in image/voxel grids. Utilized for patch and point ordering in (Lin et al., 23 Jul 2025, Li et al., 20 May 2025, Wang et al., 2024).
Consistent Traverse Serialization (CTS): Exhausts all 3! = 6 axis-permutation zigzag orderings on regularized grids; used in (Zhang et al., 2024) to enhance spatial coverage.

(b) Spectral and Graph-Based Traversals

Spectral Informed Mamba employs the spectrum of the random-walk Laplacian over patch graphs to induce isometry-invariant, manifold-aware orderings (Bahri et al., 6 Mar 2025).

Eigenvectors $v^{(k)}$ of $L_{rw} = I - D^{-1}W$ are used so that, for each mode, patches are sorted by value in both the forward and backward direction per Mamba block.

(c) Learnable/Dynamic Permutations

Point Mamba Adapter (PMA) (Zha et al., 27 May 2025) proposes a geometry-constrained gate prompt generator (G2PG) that learns spatial orderings adaptively per layer and per sample by building k-NN graphs and producing permutation indices via linear projections and argmax.

(d) Axis-wise Sorting and Group-based Ordering

Segmentation tasks benefit from axis-wise or patch-wise sortings, where points or patches are ordered along x/y/z and individual orderings are concatenated or interleaved (Lin et al., 23 Jul 2025). Hybrid Transformer-Mamba frameworks further allow intra-group Transformers with Mamba modeling at the inter-group level, prioritized by learned bi-directional importance-aware orderings (Wang et al., 2024).

3. Serialized Point Mamba Architectures

Once spatial serialization is defined, sequences are embedded and processed via parameter- and memory-efficient SSM (Mamba) blocks. Key architectural components include:

SSM Forward/Backward Scans: Both single-pass (causal) and bidirectional Mamba blocks are used to enhance global and local context capture (Liu et al., 2024, Zhang et al., 7 Jun 2025).
Selective SSM and Input Dependency: Many variants adapt the SSM parameters (transition, emission) based on the input at each token, improving capacity to model complex dependencies (Bahri et al., 6 Mar 2025, Lin et al., 23 Jul 2025).
Conditional Positional Encoding (CPE): Instead of fixed sinusoids, architectures often learn sparse 3D submanifold convolutions to inject location priors, maintaining spatial awareness under arbitrary serialization (Wang et al., 2024).
Order/Prompt Tokens: Some models prepend trainable prompts encoding the selected serialization strategy; others, as in registration (Liu et al., 16 Jun 2025), show that omitting such order indicators can improve accuracy.
Hierarchical and Hybrid Pipelines: Several works use hierarchical grouping (e.g., FPS+KNN patching, U-Net grid pooling), multi-scale serialization, and integration of Transformer modules for local or intra-group modeling, complemented by Mamba's global sequence modeling (Wang et al., 2024, Lin et al., 23 Jul 2025, Liu et al., 17 Mar 2025).

4. Applications and Downstream Tasks

Serialized Point Mamba architectures have been successfully applied to diverse geometric learning and 3D vision tasks:

Task Type	Serialization	Key Results
Object Classification	Z-order, Hilbert, Laplacian spectral	ModelNet40: 93.4% (Point Mamba), 94.5% (PointLAMA)
Part/Scene Segmentation	Hierarchical spectral, HLT, axis-sort	ShapeNetPart mIoU: 87.5% (PointLAMA), 85.9% (Spectral Mamba)
Instance Segmentation	Space-filling curves, staged	ScanNet: 76.8% mIoU, 40.0% mAP (Wang et al., 2024)
Point Cloud Registration	Z-order, Hilbert	3DMatch recall: 95.54% (Liu et al., 16 Jun 2025)
Few/Zero-shot Learning	PMA serialization	ModelNet40 5w10s: 98.8% (Zha et al., 27 May 2025)
Point Cloud Generation	Z-order/Hilbert latent sequencing	1-NNA-Abs50 EMD: 0.14%, COV: 57.90% (Liu et al., 17 Mar 2025)
Streaming Detection	Polar (sector-wise) serialization	Waymo L2 mAPH: 70.6% at 2× throughput (Zhang et al., 7 Jun 2025)

In all cases, serialization allows the SSM/Mamba backbone to operate in linear time, achieving parity or advances over prior quadratic-cost architectures (e.g., Transformer, MLP) while enabling scaling to high point counts and low-latency deployment.

5. Practical Considerations and Computational Complexity

Serialized Point Mamba designs are specifically tailored for efficiency:

Linear Complexity: Both serialization (space-filling, graph-based) and Mamba SSM blocks can be implemented in $O(N)$ or $O(M)$ where $N$ is the number of points/patches and $M$ is the sequence length, in contrast to $O(N^2)$ for attention-based models (Liu et al., 2024, Wang et al., 2024, Zha et al., 27 May 2025).
Memory Use: Only a fixed-size hidden state and kernel are stored per SSM block, yielding dramatic reduction in GPU memory at high token counts (Liu et al., 16 Jun 2025).
Staged Hierarchical Modeling: Many pipelines apply coarse-to-fine or staged SSMs (split-local/global), grid pooling, and patch-based parallelism to exploit hardware efficiently and match the multi-scale properties of 3D data.
Task-Specific Serialization: The choice of serialization (e.g., Hilbert for classification, axis-wise for segmentation), number and depth of SSM/PMLA blocks, and whether prompts or order markers are provided directly affect empirical results and should be tuned per task and dataset (Lin et al., 23 Jul 2025, Bahri et al., 6 Mar 2025, Liu et al., 16 Jun 2025).

6. Extensions: Learnable Orderings and Hybrid Models

Recent work extends static spatial serialization to dynamic or learnable strategies:

Learned Orderings: PMA and PoinTramba frameworks construct data- and task-adaptive permutations using geometric features, gate prompts, or importance scores to optimize information flow through the serialized sequence (Zha et al., 27 May 2025, Wang et al., 2024).
Hybrid Architectures: Combining Transformer modules for local context with Mamba for efficient global sequence modeling has proven effective. For example, intra-group Transformers followed by Mamba over group embeddings enable both high accuracy and tractable scaling (Wang et al., 2024).
Diffusion/Generative Modeling: Serialized Mamba backbones have been applied to efficient point cloud generation in latent space, leveraging space-filling serialization, selective SSM, and time-varying frequency-based downsampling (Liu et al., 17 Mar 2025).

7. Benchmarks and Empirical Evidence

Empirical comparisons robustly establish the effectiveness of appropriate serialization in Serialized Point Mamba:

Model / Variant	Task/Dataset	Metric/Value	Reference
Spectral Informed Mamba	ScanObjectNN	OA: 92.3%, PB-T50-RS: 87.3%	(Bahri et al., 6 Mar 2025)
Serialized Point Mamba	ScanNet (segmentation)	mIoU: 76.8%, Instance mAP: 40.0%	(Wang et al., 2024)
Point Mamba / PCM	ModelNet40	OA: 93.4%	(Liu et al., 2024, Zhang et al., 2024)
PMA	ModelNet40 5w10s	98.8% Few-shot	(Zha et al., 27 May 2025)
MT-PCR	3DMatch (registration)	RR: 95.54%	(Liu et al., 16 Jun 2025)
TFDM	ShapeNet-v2 (gen)	1-NNA-Abs50 EMD: 0.14%, COV: 57.90%	(Liu et al., 17 Mar 2025)
PHiM	Waymo (detection)	L2 mAPH: 70.6% @ 2× speed	(Zhang et al., 7 Jun 2025)

Ablation studies consistently demonstrate that spatially- and task-optimized serialization, especially those leveraging spectral/Laplacian, Hilbert/Z-order, or learned groupings, produce higher task accuracy and sample efficiency than naive or random scan, plain sortings, or non-serialized SSM (Bahri et al., 6 Mar 2025, Liu et al., 16 Jun 2025, Lin et al., 23 Jul 2025). Masked autoencoding, token reordering/restoration, and localized attention blocks further enhance performance in self-supervised and semi-supervised settings.

In sum, Serialized Point Mamba encompasses the methodology of transforming unordered/irregular geometric data into sequences amenable for linear-complexity SSM (Mamba) modeling, with the integrity of spatial relationships maintained by principled serialization—be it fractal, spectral, hierarchical, or dynamically learned. This design paradigm achieves state-of-the-art results in diverse 3D learning tasks, and serves as a reference standard for efficient and scalable geometric deep learning (Liu et al., 2024, Bahri et al., 6 Mar 2025, Zha et al., 27 May 2025, Liu et al., 16 Jun 2025, Wang et al., 2024).