Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Nd-BiMamba2: A Unified Bidirectional Architecture for Multi-Dimensional Data Processing

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A single Mamba2-based architecture, Nd-BiMamba2, processes 1D, 2D, and 3D data with bidirectional context and deploys across CPU, GPU, and mobile.

desk verdict A clean, honest engineering wrapper for bidirectional Mamba2 across 1D/2D/3D with shipped code and portability tests—but the claim that it improves feature representation rests on a category error: FLOPs and timing are treated as evidence of accuracy, and no task-level experiment appears anywhere. read the letter →

arxiv 2411.15380 v1 pith:Y64LHX72 submitted 2024-11-22 cs.LG cs.AI

Hao Liu This is my paper · ORCID
classification cs.LGcs.AI
keywords Mamba2statespacemodelsbidirectionalmodelingmulti-dimensionaldataprocessingadaptivepaddingONNXTorchScriptcross-platformdeployment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Nd-BiMamba2 proposes a unified architecture built on the Mamba2 state-space module for processing 1D time series, 2D images, and 3D volumes with the same code path. The design runs a forward Mamba2 pass and a backward Mamba2 pass over the reversed input, adds the two feature maps, and uses padding sizes adapted to the input dimension so that shapes stay regular. The paper argues this captures bidirectional context in higher-dimensional data while keeping the linear-time efficiency of state-space models and avoiding the quadratic cost of attention. If correct, a single model family could replace separate CNN, RNN, and Transformer designs for different data dimensions and be exported to ONNX and TorchScript for CPU, GPU, and mobile deployment.

What carries the argument

The central object is the bidirectional Mamba2 block: a forward Mamba2 pass over the flattened input and a second pass over the reversed input, with the two hidden representations added and then reordered. Mamba2 is a selective state-space layer that models sequences in linear time; running it in both directions is what supplies bidirectional context. The adaptive padding step, with padding size computed from kernel size and stride per dimension, keeps tensors aligned to fixed multiples so the same convolution and state-space code can handle 1D, 2D, and 3D inputs. The flatten-and-reshape step is the load-bearing device that lets one recurrence sweep across all dimensions.

What would settle it

Train Nd-BiMamba2 and a standard 2D-specific model on the same image classification benchmark, such as CIFAR-10, with matched budgets, and also train Nd-BiMamba2 with row-major versus randomly shuffled flattening orders; if accuracy is similar under shuffling, or far below the 2D-specific baseline, the flatten-then-scan design is not capturing spatial geometry.

Watch

Extended reading notes

Core claim

The central claim is that the Mamba2 module can be extended into a general bidirectional engine for data of any rank by uniformly representing inputs as a tensor with dummy dimensions for missing axes, flattening the tensor to a sequence, running separate forward and backward Mamba2 paths, and fusing them by addition. Nd-BiMamba2 wraps this with dimension-adaptive convolution kernels—$(k,1,1)$ for 1D, $(k_1,k_2,1)$ for 2D, $(k_1,k_2,k_3)$ for 3D—and an adaptive padding rule that aligns each dimension to a multiple of the model's block size (4 for 1D and 3D, 8 for 2D). The output is trimmed back to the original shape, making the block a drop-in layer for inputs of different dimensionalities. The paper's evidence is computational and portability focused: it reports FLOPs, latency, and parameter counts on CPU, GPU, and mobile, with bidirectional processing roughly doubling FLOPs and parameters while the layer count remains fixed.

Load-bearing premise

The load-bearing premise is that flattening 2D and 3D tensors into 1D sequences, in the order the code chooses, preserves enough spatial structure for the state-space model to learn useful features.

Editorial extensions

If this is right

  • One architecture can serve the three most common data ranks, so a project spanning time series, images, and volumes no longer needs separate model families.
  • The block preserves input and output shapes, so it can slot into existing networks at any layer without dimension-specific rewiring.
  • Because the model exports to ONNX and TorchScript, the same trained weights can run in production on CPUs, GPUs, and mobile devices.
  • Bidirectional processing increases FLOPs and computation time about twofold, but the added parameter count is independent of input size.
  • Adaptive padding reduces memory waste compared with fixed equal padding, particularly for non-square 2D and non-cubic 3D inputs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The flattening order, whether row-major, snake, or Hilbert-style, is an implicit design choice that could matter more than the bidirectional fusion itself, and the paper does not test it.
  • The same forward-plus-backward-then-add recipe could be applied to other linear-time sequence models, such as linear attention, to isolate whether the benefit comes from bidirectionality or from Mamba2's state compression.
  • If the architecture's generality holds, dimensionality could become a hyperparameter rather than an architectural commitment, allowing one pretrained module to be adapted across audio, image, and volumetric tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes Nd-BiMamba2, a unified architecture based on the Mamba2 state-space model, intended to process 1D, 2D, and 3D data by flattening inputs, running forward and backward Mamba2 modules, fusing their outputs, and undoing padding. The authors report computational cost (FLOPs, time, parameters) across 1D/2D/3D sizes, demonstrate ONNX/TorchScript export, and test on CPU/GPU; code is open-source. The central claim is that this bidirectional design captures multi-dimensional information efficiently.

Significance. If the architecture were shown to improve task-level accuracy on 1D/2D/3D benchmarks, it would be a useful contribution given its modular design and cross-platform deployment. The open-source code and export support are strengths for reproducibility. However, the paper as written provides no accuracy experiments; the claimed feature-representation improvement rests solely on increased FLOPs and runtime, which by themselves do not establish utility. The related work also incorrectly states that existing Mamba modules lack bidirectionality while citing bidirectional variants. Thus the scientific contribution is currently limited to a deployment demo, not a validated model.

major comments (4)
  1. [V-B, Table III] The claim that 'comparative experiments... validated the improvement in feature representation by bidirectional modeling' is unsupported: Table III reports only FLOPs, time, and parameter counts, with no accuracy measure. Increased FLOPs, time, and an 89% parameter increase (150.8k -> 285.21k) are costs, not evidence of improved feature quality. Please add task-level evaluations (e.g., classification or regression on 1D, 2D, and 3D datasets) comparing Nd-BiMamba2 against unidirectional Mamba2 and existing bidirectional Mamba baselines, with standard metrics and error bars.
  2. [V-C, Table IV] The statement that adaptive padding 'maintains high accuracy across different input sizes' is not supported by Table IV, which lists only input dimensions, auto-padded sizes, and a boolean indicator of whether the padding matches Mamba2's equal padding. No accuracy is reported anywhere in the paper. Please either remove accuracy claims or provide the corresponding measurements.
  3. [II-D] The related-work section asserts that 'existing Mamba modules primarily focus on unidirectional feature modeling' and lack bidirectional modeling, yet the same section cites [20] (VSSD, non-causal state space duality) and [22] (Vision Mamba, bidirectional state space model). This internal contradiction overstates the novelty of Nd-BiMamba2. The manuscript should accurately position its contribution relative to these existing bidirectional Mamba works and empirically compare against them.
  4. [IV-B-1 and Algorithm 1 Step 1] The design flattens 2D/3D tensors into a 1D sequence (Section IV-B-1, Algorithm 1 Step 1), which determines the spatial context available to the state-space model. The paper does not justify that this flattening preserves useful spatial structure, nor does it test alternative scan orders. Without accuracy experiments, the reader cannot tell whether the model exploits multi-dimensional structure or merely discards it. Please include ablations over scan orders and comparisons with 2D/3D-aware Mamba methods.
minor comments (5)
  1. [III] Section III ('Algorithm Design') appears to be empty; the text jumps from Section II directly to Section IV. Please either merge the sections or add the missing content.
  2. [Table III] The unit 'GMac' in the FLOPs column denotes multiply-accumulate operations, not floating-point operations; please use a consistent metric (e.g., MACs or FLOPs).
  3. [V-B] The text says the parameter-count growth is 'relatively small,' but the number of parameters nearly doubles from 150.8k to 285.21k (about 89% increase); please revise this characterization.
  4. [Equation (7)] The padding formula as typeset is ambiguous: it should be \max(0, \lceil ((D_i - 1)s_i + k_i - 1)/2 \rceil) or similar. Please clarify the numerator and parentheses.
  5. [General] There are several typos, including 'nd-BiMamba2' (lowercase) in Section V-D, 'B IMAMBA2' in the index terms, and the unrendered section headers. A thorough proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the architecture is described constructively, and no fitted input is relabeled as a prediction.

full rationale

Nd-BiMamba2's core algorithm (Algorithm 1) is a constructive pipeline: pad, reshape, channel-map, run forward and backward Mamba2, sum, map back, and trim. The paper reports no parameter fitted to a target and no result whose definition contains its conclusion. Equation (7) is a standard padding calculation and is not used to fabricate an empirical outcome. Tables III and IV report FLOPs, timing, and shape data, not accuracy; the text's claims that bidirectional modeling improves expressive power and that adaptive padding maintains high accuracy are unsupported by accuracy experiments, but unsupported empirical strength is a correctness risk, not circularity. There are no self-citations, imported uniqueness theorems, or ansatz-smuggling citations; the dependence on the external Mamba2 module is openly stated, and the contribution is an engineering composition of that module. The bidirectional fusion is a straightforward sum of forward and backward features, and the padding rule follows from the kernel and stride definitions rather than from the results it is used to discuss. Therefore the derivation chain is self-contained in the sense relevant to circularity, and the score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper rests on the prior Mamba2 architecture and on the assumption that flattening preserves structure. No numbers are fitted to data; the model hyperparameters reported in experiments are configuration choices, not free parameters. No new entities are invented.

assumptions (4)
  • domain assumption Mamba2 provides an efficient linear-time sequence model
    The entire architecture wraps the Mamba2 module from Ref [19]; the paper assumes its efficiency and correctness without re-derivation.
  • domain assumption Flattening 2D/3D tensors to a 2D matrix preserves usable spatial structure for SSM scanning
    Section IV-B-1 'Dimension Adjustment' and Appendix Algorithm 1 Step 1 reshape the data; the paper never validates this against task accuracy.
  • domain assumption Forward and backward passes with additive fusion capture bidirectional information
    Appendix Algorithm 1 Step 2 flips the input and adds hidden states; the paper asserts this is an innovative bidirectional processing mechanism without proof or task evidence.
  • standard math Standard padding formula for convolutional output size
    Equation (7) is a standard calculation for padding size given kernel size and stride; it is not specific to this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nd-BiMamba2: A Unified Bidirectional Architecture for Multi-Dimensional Data Processing." pith.science (2026). https://pith.science/paper/Y64LHX72

@misc{pith2026241115380,
  author       = {Pith},
  title        = {Pith review of: Nd-BiMamba2: A Unified Bidirectional Architecture for Multi-Dimensional Data Processing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y64LHX72}},
  note         = {Machine review of arXiv:2411.15380}
}
read the original abstract

Deep learning models often require specially designed architectures to process data of different dimensions, such as 1D time series, 2D images, and 3D volumetric data. Existing bidirectional models mainly focus on sequential data, making it difficult to scale effectively to higher dimensions. To address this issue, we propose a novel multi-dimensional bidirectional neural network architecture, named Nd-BiMamba2, which efficiently handles 1D, 2D, and 3D data. Nd-BiMamba2 is based on the Mamba2 module and introduces innovative bidirectional processing mechanisms and adaptive padding strategies to capture bidirectional information in multi-dimensional data while maintaining computational efficiency. Unlike existing methods that require designing specific architectures for different dimensional data, Nd-BiMamba2 adopts a unified architecture with a modular design, simplifying development and maintenance costs. To verify the portability and flexibility of Nd-BiMamba2, we successfully exported it to ONNX and TorchScript and tested it on different hardware platforms (e.g., CPU, GPU, and mobile devices). Experimental results show that Nd-BiMamba2 runs efficiently on multiple platforms, demonstrating its potential in practical applications. The code is open-source: https://github.com/Human9000/nd-Mamba2-torch

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 11 canonical work pages

  1. [20]

    Vssd: Vision mamba with non- causal state space duality,

    Y . Shi, M. Dong, M. Li, and C. Xu, “Vssd: Vision mamba with non- causal state space duality,” arXiv preprint arXiv:2407.18559 , 2024

    arXiv 2024

  2. [22]

    Vi sion mamba: Efficient visual representation learning with bidir ectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vi sion mamba: Efficient visual representation learning with bidir ectional state space model,” arXiv preprint arXiv:2401.09417 , 2024

    arXiv 2024

  3. [1]

    From cn n to transformer: A review of medical image segmentation models ,

    W. Y ao, J. Bai, W. Liao, Y . Chen, M. Liu, and Y . Xie, “From cn n to transformer: A review of medical image segmentation models ,” Journal of Imaging Informatics in Medicine , pp. 1–19, 2024

    work page 2024

  4. [2]

    Rnn-lstm: From appl ications to modeling techniques and beyond—systematic review,

    S. M. Al-Selwi, M. F. Hassan, S. J. Abdulkadir, A. Muneer, E. H. Sumiea, A. Alqushaibi, and M. G. Ragab, “Rnn-lstm: From appl ications to modeling techniques and beyond—systematic review,” Journal of King Saud University-Computer and Information Sciences , p. 102068, 2024

    work page 2024

  5. [3]

    Stacked bidirectional a nd unidirec- tional lstm recurrent neural network for forecasting netwo rk-wide traffic state with missing values,

    Z. Cui, R. Ke, Z. Pu, and Y . Wang, “Stacked bidirectional a nd unidirec- tional lstm recurrent neural network for forecasting netwo rk-wide traffic state with missing values,” Transportation Research Part C: Emerging Technologies, vol. 118, p. 102674, 2020

    work page 2020

  6. [4]

    Attention is all you need,

    A. V aswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  7. [5]

    Mamba: Linear-time sequence modeling w ith selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling w ith selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    arXiv 2023

  8. [6]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, and P . Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

    work page 1998

Show all 22 references
  1. [7]

    Deep residual learnin g for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learnin g for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  2. [8]

    U-net: Convolu tional networks for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-net: Convolu tional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th intern ational con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III ...

    2015

  3. [9]

    Fundamentals of recurrent neural netw ork (rnn) and long short-term memory (lstm) network,

    A. Sherstinsky, “Fundamentals of recurrent neural netw ork (rnn) and long short-term memory (lstm) network,” Physica D: Nonlinear Phe- nomena, vol. 404, p. 132306, 2020

    2020

  4. [10]

    Gate-variants of gated recurren t unit (gru) neural networks,

    R. Dey and F. M. Salem, “Gate-variants of gated recurren t unit (gru) neural networks,” in 2017 IEEE 60th international midwest symposium on circuits and systems (MWSCAS) . IEEE, 2017, pp. 1597–1600

    2017

  5. [11]

    Bidirectional lo ng short- term memory networks for relation classification,

    S. Zhang, D. Zheng, X. Hu, and M. Y ang, “Bidirectional lo ng short- term memory networks for relation classification,” in Proceedings of the 29th Pacific Asia conference on language, information and co mputation, 2015, pp. 73–78

    2015

  6. [12]

    Wavenet: A generative model for raw audio,

    A. V an Den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Viny als, A. Graves, N. Kalchbrenner, A. Senior, K. Kavukcuoglu et al. , “Wavenet: A generative model for raw audio,” arXiv preprint arXiv:1609.03499, vol. 12, 2016

    2016 arXiv

  7. [13]

    Bert: Pre-training of deep bidirectional t ransformers for language understanding,

    J. Devlin, “Bert: Pre-training of deep bidirectional t ransformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

    2018 arXiv

  8. [14]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenbor n, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gel ly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Onlin e]. Available: https://a...

    2021 arXiv

  9. [15]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B . Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” 2021. [Online]. Available: https://arxiv.org/abs/2103.14030

    2021 arXiv

  10. [16]

    Swin transformer v2: Scaling up capacity and resolution,

    Z. Liu, H. Hu, Y . Lin, Z. Y ao, Z. Xie, Y . Wei, J. Ning, Y . Cao , Z. Zhang, L. Dong et al., “Swin transformer v2: Scaling up capacity and resolution,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 12 009–12 019

    2022

  11. [17]

    Linforme r: Self-attention with linear complexity,

    S. Wang, B. Z. Li, M. Khabsa, H. Fang, and H. Ma, “Linforme r: Self-attention with linear complexity,” arXiv preprint arXiv:2006.04768, 2020

    2006 arXiv

  12. [18]

    Rethinking attention with performers,

    K. Choromanski, V . Likhosherstov, D. Dohan, X. Song, A. Gane, T. Sar- los, P . Hawkins, J. Davis, A. Mohiuddin, L. Kaiser et al. , “Rethinking attention with performers,” arXiv preprint arXiv:2009.14794 , 2020

    2009 arXiv

  13. [19]

    Transformers are ssms: Generalized mo dels and ef- ficient algorithms through structured state space duality,

    T. Dao and A. Gu, “Transformers are ssms: Generalized mo dels and ef- ficient algorithms through structured state space duality, ” arXiv preprint arXiv:2405.21060, 2024

    2024 arXiv

  14. [21]

    Mamba- in-mamba: Centralized mamba-cross-scan in tokenized mamb a model for hyperspectral image classification,

    W. Zhou, S.-i. Kamata, H. Wang, M. S. Wong, and H. C. Hou, “ Mamba- in-mamba: Centralized mamba-cross-scan in tokenized mamb a model for hyperspectral image classification,” Neurocomputing, p. 128751, 2024

    2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.