Pith. sign in

REVIEW 3 major objections 5 minor 27 references

MambaNeXt-YOLO: A Hybrid State Space Model for Real-time Object Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Hybridizing a convolutional local branch with a Mamba state-space global branch, MambaNeXt-YOLO reports 66.6% mAP on PASCAL VOC and 27.8% on DOTA v1.5 without pre-training, with 31.9 FPS on an edge device.

desk verdict Competent engineering, but the core Mamba claim rests on a recurrence that is not the standard Mamba update, and the efficiency headline is contradicted by the paper's own speed table. read the letter →

arxiv 2506.03654 v3 pith:UBZUZP3A submitted 2025-06-04 cs.CV cs.AI

Xiaochun Lei , Siqi Wu , Weilin Wu , Zetao Jiang This is my paper
classification cs.CVcs.AI
keywords real-timeobjectdetectionstatespacemodelsMambaNeXt-YOLOfeaturepyramidnetworkedgedeploymenthybridCNN-SSMmulti-scale
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

The paper sets out to show that a real-time object detector can get long-range context from a linear state-space model without giving up the speed that makes YOLO-style detectors practical. It proposes MambaNeXt-YOLO, whose backbone builds each stage from a hybrid MambaNeXt block: a ConvNeXt local branch, a Mamba global branch, and a ResGate adaptive fusion, followed by a multi-branch asymmetric feature pyramid neck. Trained from scratch on PASCAL VOC and DOTA v1.5, the model reports 66.6% mAP on VOC, above the closest Mamba-based baseline, Mamba-YOLO-T, at 66.0%, and 27.8% mAP on DOTA v1.5, with 31.9 FPS on a Jetson Orin NX. If those numbers hold, hybrid CNN-state-space design becomes a credible recipe for edge object detection.

What carries the argument

The central object is the MambaNeXt block, a three-part module placed throughout the backbone and neck. Its local branch is a ConvNeXt-style depthwise/pointwise convolution path; its global branch is an SS2D (two-dimensional selective-scan) Mamba core that projects each token $x_t$ into parameters $A_t$, $B_t$, $\Delta_t$ and updates a hidden state by $h_{t+1}=e^{-\Delta_t}\odot h_t+A_t+B_t\odot h_t$ before reshaping back to a feature map; its ResGate path generates two parallel projections $U$ and $V$, applies a depthwise GELU-modulated tensor, and gates it with $V$, adding residuals. The block is inserted into the MAFPN neck, which provides bidirectional top-down and bottom-up pathways and replaces max-pooling with stride-2 convolutions for downsampling.

What would settle it

Inspect the released implementation's forward pass and check whether the hidden state follows Eq. (10), $h_{t+1}=e^{-\Delta_t}\odot h_t+A_t+B_t\odot h_t$, or the standard Mamba update $h_{t+1}=\exp(\Delta A) h_t + \Delta B x_t$; a match with one form but not the other would settle what the 'Mamba' label in this paper refers to.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that combining convolutional local feature extraction with a linear state-space global model inside one block, and feeding the result into a multi-branch asymmetric fusion neck, yields a detector that is more accurate than the YOLO and SSM-based baselines it compares against while remaining fast enough for edge GPUs. The reported evidence is 66.6% mAP on PASCAL VOC and 27.8% on DOTA v1.5 without any pre-training, outperforming Mamba-YOLO-T (66.0% and 25.7%) and several YOLO variants in the same tables. The design's two named contributions are the MambaNeXt block, which interleaves ConvNeXt, SS2D Mamba, and ResGate gating, and the MAFPN neck, whose stride-2 convolution downsampling the ablation credits with raising mAP from 63.2% to 66.0%.

Load-bearing premise

The load-bearing premise is that the recurrence written as Eq. (10) is actually what the implemented Mamba module computes, since the paper supplies no code or formal specification and Eq. (10) differs from the standard Mamba update.

Editorial extensions

If this is right

  • If the reported numbers are correct, a detector can absorb a linear-complexity state-space global branch and still run at 31.9 FPS on a Jetson Orin NX with TensorRT FP16, so SSM-based global context is edge-deployable.
  • Stride-2 convolutions in the neck beat max-pooling downsampling by 2.8 mAP in the paper's ablation, making learnable downsampling a transferable design choice.
  • Training from scratch for 1,000 epochs reaches 66.6% mAP on VOC and 27.8% on DOTA v1.5, which the paper presents as evidence that hybrid CNN-SSM backbones need no pre-training to compete with YOLO variants.
  • The state dimension and SSM ratio matter: raising the state from 8 to 32 lifts VOC mAP from 65.9% to 66.6%, and increasing the SSM ratio beyond 2 hurts accuracy, giving concrete tuning guidance for similar hybrids.

Reading between the lines

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

  • The paper does not isolate the block's contribution, since it changes both the backbone block and the neck relative to Mamba-YOLO-T; swapping only one component at a time would be a cleaner test.
  • The paper does not frame it this way, but its own Table 7 shows MambaNeXt-YOLO is slower than several YOLO baselines on the same edge devices, so the contribution is better read as accuracy under an SSM budget than as a speed-up.
  • The same three-part block could in principle be dropped into other detectors; if its DOTA improvement transfers, it would offer a general way to add global context to small-object aerial detection.
  • The full pipeline is evaluated only on VOC-style and DOTA protocols with 1,000 epochs, so re-running on a standard COCO schedule would be the natural next test of the claim's generality.
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

3 major / 5 minor

Summary. The paper proposes MambaNeXt-YOLO, a real-time object detector whose backbone is built from hybrid blocks that combine a ConvNeXt-style local branch, an SS2D Mamba block for global context, and a ResGate adaptive fusion. The detector also adopts the MAFPN neck from MHAF-YOLO. Experiments on PASCAL VOC and DOTA v1.5 report 66.6% mAP and 27.8% mAP, respectively, with inference speeds on an RTX 3090 and on Jetson Orin NX / Xavier NX edge devices. The paper claims that this design balances accuracy and efficiency, supports edge deployment, and outperforms prior real-time detectors without pre-training.

Significance. If the claims are correct, the paper would contribute a lightweight hybrid CNN-SSM detector with competitive accuracy on two benchmarks and plausible edge deployment. The proposed MambaNeXt block is a reasonable architectural idea, and the inclusion of Mamba-YOLO and VMamba baselines gives a relevant comparison context. However, the paper does not ship code or a machine-checked formal specification, the speed advantage is contradicted by the paper's own Table 7, the accuracy gains over the closest Mamba-based baseline are small (0.6 and 2.1 mAP) with no error bars, and the central formal description of the Mamba recurrence (Eq. 10) does not match the standard Mamba/SS2D update. These issues place the main contribution on fragile ground.

major comments (3)
  1. [§3.3, Eq. (10)] The recurrence h_{t+1} = e^{-Δ_t} ⊙ h_t + A_t + B_t ⊙ h_t is not the standard Mamba/SS2D selective SSM update, which is h_{t+1} = exp(Δ_t A) h_t + Δ_t B_t x_t with output y_t = C_t h_t. The paper's equation omits the explicit input term Δ_t B_t x_t and the output projection C_t h_t, and the text does not describe the four-directional scanning used by SS2D. Yet §4.1 states that the MambaNeXt module is configured "including its core SS2D components." This is an internal inconsistency: either the implementation is not described by Eq. (10), or it is not an SS2D/Mamba block. Without code or a corrected formal specification, the paper's central claim of a hybrid CNN-Mamba detector is unsupported.
  2. [Abstract and Table 7] The paper's third contribution is "Edge-focused Efficiency," and the abstract claims "balanced accuracy and efficiency." However, Table 7 shows MambaNeXt-YOLO achieves 31.9 FPS on Jetson Orin, 19.5 FPS on Xavier NX, and 34.6 FPS on RTX 3090, which is slower than YOLOv8-S (67.5, 25.5, 50.3), YOLO11-S (66.2, 26.0, 51.3), and Mamba-YOLO-T (34.3, 20.9, 45.2). The efficiency claim is contradicted by the paper's own measurements, and this undermines a stated contribution.
  3. [§4.1 and Tables 1-2] Baseline training conditions are not described: it is unclear whether the YOLO and Mamba-YOLO baselines were trained from scratch with the same number of epochs, batch size, input resolution, and optimizer settings, or whether their published weights were used. The gains over the closest Mamba-based baseline, Mamba-YOLO-T, are 0.6 mAP on VOC and 2.1 mAP on DOTA, with no error bars or significance tests. The hyperparameters in Tables 5 and 6 are tuned on the validation set, and the reported 66.6% mAP is the peak of this sweep; this fragility should be acknowledged.
minor comments (5)
  1. [§4.1] The paragraph beginning "As shown in Table 1 and Table 2..." is duplicated verbatim immediately after the implementation details paragraph.
  2. [§3.1 vs. Abstract] The abstract defines MAFPN as "Multi-branch Asymmetric Fusion Pyramid Network" while §3.1 calls it "Multi-Asymmetric Fusion Pyramid Network"; the acronym should be defined consistently.
  3. [Tables 4-6] The exact configuration of the final model (block order, dstate, SSM ratio, MLP ratio, ConvNeXt kernel size, layer scale) is never assembled in one place; the reader must infer it from the peaks of separate ablation tables, which is error-prone.
  4. [§3.3] The dimensions of Fscan and of the generated parameters A_t, B_t, and Δ_t are not specified; the text states "reshape Fscan ∈ R^{C×H×W}" but does not explain the channel dimension after the linear projection in Eq. (6).
  5. [Tables 1-2 vs. Table 7] Tables 1 and 2 do not report FPS or latency, so the main accuracy comparison is not presented together with the speed comparison from Table 7, making the speed-accuracy trade-off difficult to evaluate.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reported mAP/FPS values are empirical benchmarks against external baselines; the validation/test overlap is an evaluation concern rather than a circular derivation.

full rationale

The paper does not derive a target quantity from an input quantity; its contributions are architectural (MambaNeXt block, MAFPN neck, detection head) and are validated by training on PASCAL VOC and DOTA v1.5 and comparing against external YOLO and Mamba-YOLO baselines. The 66.6% mAP on PASCAL VOC and 27.8% on DOTA v1.5 are end-to-end training results, not quantities computable from the hyperparameters or from any fitted parameter by construction. Hyperparameter choices (dstate, SSM ratio, MLP ratio, ConvNeXt settings) are ablated empirically and are not renamed predictions of the headline numbers. The one caveat is that Section 4.1 states 'the VOC 2007 test set was used for both validation and testing', and Table 5 selects hyperparameters by validation mAP; this is a test-set leakage/selection concern that inflates the reported result, but it is not a circularity because the reported mAP still requires training and measuring the stated model. There is no load-bearing self-citation: references to VMamba, Mamba-YOLO, and MHAF-YOLO are external prior work, and no uniqueness theorem is invoked to force a choice. The discrepancy between Eq. (10) and the standard SS2D recurrence (omitting the explicit input term and the output projection) is a correctness and verifiability problem for the 'Mamba' attribution, not a circular derivation, and is left for the correctness pass.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The central empirical claim depends on hyperparameters tuned on the validation set and on several unverified protocol assumptions. No new theoretical entities are introduced.

free parameters (7)
  • State dimension (dstate) = 32
    Tuned in Table 5; chosen because it gave the best mAP (66.6%). This is a free parameter of the Mamba recurrence.
  • SSM expansion ratio = 2
    Tuned in Table 5 as the best balance of accuracy and cost.
  • MLP expansion ratio = 4
    Tuned in Table 5; higher values added parameters without improving mAP.
  • ConvNeXt kernel size = 7
    Tuned in Table 6; larger kernel gave the highest mAP.
  • Layer scale initialization = 1e-6
    Tuned in Table 6 as the best initialization value.
  • Feature width multiplier = 4x
    Tuned in Table 6; 4x width gave best accuracy without extra overhead.
  • Input resolution = 640x640
    Hand-chosen operating point; affects all accuracy and speed numbers.
assumptions (3)
  • domain assumption The SS2D Mamba block used in the implementation behaves as a standard selective state space model.
    Section 3.3, Eq. (7)-(10), describes an update scheme that does not match the standard Mamba recurrence; the paper provides no code or formal description bridging this gap.
  • domain assumption Baseline detectors in Tables 1, 2, and 7 were evaluated under a comparable training and benchmarking protocol.
    Section 4.1 gives implementation details only for MambaNeXt-YOLO; no pretraining status, epochs, or TensorRT settings are given for the baselines.
  • domain assumption 1000 epochs of from-scratch training is a fair comparison regime for all methods.
    The paper claims a strong result without pretraining but does not report whether YOLO baselines also trained from scratch for 1000 epochs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MambaNeXt-YOLO: A Hybrid State Space Model for Real-time Object Detection." pith.science (2026). https://pith.science/paper/UBZUZP3A

@misc{pith2026250603654,
  author       = {Pith},
  title        = {Pith review of: MambaNeXt-YOLO: A Hybrid State Space Model for Real-time Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UBZUZP3A}},
  note         = {Machine review of arXiv:2506.03654}
}
read the original abstract

Real-time object detection is a fundamental but challenging task in computer vision, particularly when computational resources are limited. Although YOLO-series models have set strong benchmarks by balancing speed and accuracy, the increasing need for richer global context modeling has led to the use of Transformer-based architectures. Nevertheless, Transformers have high computational complexity because of their self-attention mechanism, which limits their practicality for real-time and edge deployments. To overcome these challenges, recent developments in linear state space models, such as Mamba, provide a promising alternative by enabling efficient sequence modeling with linear complexity. Building on this insight, we propose MambaNeXt-YOLO, a novel object detection framework that balances accuracy and efficiency through three key contributions: (1) MambaNeXt Block: a hybrid design that integrates CNNs with Mamba to effectively capture both local features and long-range dependencies; (2) Multi-branch Asymmetric Fusion Pyramid Network (MAFPN): an enhanced feature pyramid architecture that improves multi-scale object detection across various object sizes; and (3) Edge-focused Efficiency: our method achieved 66.6% mAP at 31.9 FPS on the PASCAL VOC dataset without any pre-training and supports deployment on edge devices such as the NVIDIA Jetson Xavier NX and Orin NX.

Figures

Figures reproduced from arXiv: 2506.03654 by the authors.

Figure 1
Figure 1. Illustration of the overall architecture of MambaNeXt-YOLO. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Illustration of the Vision Clue Merge Block. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. Illustration of the MambaNeXt block architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 13 canonical work pages

  1. [1]

    arXiv preprint arXiv:2004.10934 (2020)

    Bochkovskiy, A., Wang, C.Y ., Liao, H.Y .M.: Yolov4: Op- timal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020)

    arXiv 2020

  2. [2]

    Everingham, M., Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Int. J. Comput. Vision 88(2), 303–338 (Jun 2010). https: //doi.org/10.1007/s11263-009-0275-4, https:// doi.org/10.1007/s11263-009-0275-4

    doi:10.1007/s11263-009-0275-4 2010

  3. [3]

    arXiv preprint arXiv:2312.00752 (2023)

    Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)

    arXiv 2023

  4. [4]

    arXiv preprint arXiv:2111.00396 (2021)

    Gu, A., Goel, K., R ´e, C.: E fficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021)

    arXiv 2021

  5. [5]

    arXiv preprint arXiv:2403.09338 (2024)

    Huang, T., Pei, X., You, S., Wang, F., Qian, C., Xu, C.: Local- mamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338 (2024)

    arXiv 2024

  6. [6]

    https://doi.org/10.5281/zenodo.3908559, https://github

    Jocher, G.: YOLOv5 by Ultralytics (May 2020). https://doi.org/10.5281/zenodo.3908559, https://github. com/ultralytics/yolov5 6 Table 7: Inference speed comparison on NVIDIA devices. Speed is measured in frames per second (FPS). Method FPSOrin FPSNX FPSRTX3090 YOLOv8-S 67.5 25.5 50.3 YOLOv9-T 27.1 18.7 36.8 YOLOv10-S 38.9 20.2 39.8 YOLO11-S 66.2 26.0 51.3 ...

    doi:10.5281/zenodo.3908559 2020

  7. [7]

    Jocher, G., Qiu, J., Chaurasia, A.: Ultralytics YOLO (Jan 2023), https://github.com/ultralytics/ultralytics

    work page 2023

  8. [8]

    arXiv preprint arXiv:2209.02976 (2022)

    Li, C., Li, L., Jiang, H., Weng, K., Geng, Y ., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022)

    arXiv 2022

Show all 27 references
  1. [9]

    Advances in Neural Information Processing Sys- tems 35, 12934–12949 (2022)

    Li, Y ., Yuan, G., Wen, Y ., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y ., Ren, J.: Efficientformer: Vision transformers at mo- bilenet speed. Advances in Neural Information Processing Sys- tems 35, 12934–12949 (2022)

    2022

  2. [10]

    arXiv preprint arXiv:2401.10166 (2024)

    Liu, Y ., Tian, Y ., Zhao, Y ., Yu, H., Xie, L., Wang, Y ., Ye, Q., Liu, Y .: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024)

    2024 arXiv

  3. [11]

    arXiv preprint arXiv:2110.02178 (2021)

    Mehta, S., Rastegari, M.: Mobilevit: light-weight, general- purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178 (2021)

    2021 arXiv

  4. [12]

    In: Eu- ropean conference on computer vision

    Pan, J., Bulat, A., Tan, F., Zhu, X., Dudziak, L., Li, H., Tz- imiropoulos, G., Martinez, B.: Edgevits: Competing light- weight cnns on mobile devices with vision transformers. In: Eu- ropean conference on computer vision. pp. 294–311. Springer (2022)

    2022

  5. [13]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR) (June 2016)

    Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR) (June 2016)

    2016

  6. [14]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

    Redmon, J., Farhadi, A.: Yolo9000: Better, faster, stronger. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

    2017

  7. [15]

    arXiv preprint arXiv:1804.02767 (2018)

    Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)

    2018 arXiv

  8. [16]

    arXiv preprint arXiv:2502.12524 (2025)

    Tian, Y ., Ye, Q., Doermann, D.: Yolov12: Attention-centric real-time object detectors. arXiv preprint arXiv:2502.12524 (2025)

    2025 arXiv

  9. [17]

    Advances in neural information processing systems 30 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)

    2017

  10. [18]

    Advances in Neural Information Processing Systems 37, 107984–108011 (2024)

    Wang, A., Chen, H., Liu, L., Chen, K., Lin, Z., Han, J., et al.: Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems 37, 107984–108011 (2024)

    2024

  11. [19]

    In: Proceedings of the IEEE /CVF conference on com- puter vision and pattern recognition

    Wang, C.Y ., Bochkovskiy, A., Liao, H.Y .M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object de- tectors. In: Proceedings of the IEEE /CVF conference on com- puter vision and pattern recognition. pp. 7464–7475 (2023)

    2023

  12. [20]

    In: European conference on computer vision

    Wang, C.Y ., Yeh, I.H., Mark Liao, H.Y .: Yolov9: Learning what you want to learn using programmable gradient information. In: European conference on computer vision. pp. 1–21. Springer (2024)

    2024

  13. [21]

    Wang, Z., Li, C., Xu, H., Zhu, X.: Mamba yolo: Ssms-based yolo for object detection (2024), https://arxiv.org/abs/ 2406.05835

    2024 arXiv

  14. [22]

    In: The IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) (June 2018)

    Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: Dota: A large-scale dataset for object detection in aerial images. In: The IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) (June 2018)

    2018

  15. [23]

    CAAI transactions on intelligence technology 8(2), 319–330 (2023)

    Yan, M., Lou, X., Chan, C.A., Wang, Y ., Jiang, W.: A semantic and emotion-based dual latent variable generation model for a dialogue system. CAAI transactions on intelligence technology 8(2), 319–330 (2023)

    2023

  16. [24]

    IEEE Transactions on Vehicular Technology 73(4), 5647–5658 (2023)

    Yan, M., Xiong, R., Wang, Y ., Li, C.: Edge computing task of- floading optimization for a uav-assisted internet of vehicles via deep reinforcement learning. IEEE Transactions on Vehicular Technology 73(4), 5647–5658 (2023)

    2023

  17. [25]

    IEEE Sensors Journal (2024)

    Yan, M., Zhang, L., Jiang, W., Chan, C.A., Gygax, A.F., Nir- malathas, A.: Energy consumption modeling and optimization of uav-assisted mec networks using deep reinforcement learn- ing. IEEE Sensors Journal (2024)

    2024

  18. [26]

    arXiv preprint arXiv:2502.04656 (2025)

    Yang, Z., Guan, Q., Yu, Z., Xu, X., Long, H., Lian, S., Hu, H., Tang, Y .: Mhaf-yolo: Multi-branch heterogeneous auxil- iary fusion yolo for accurate object detection. arXiv preprint arXiv:2502.04656 (2025)

    2025 arXiv

  19. [27]

    In: Forty-first International Confer- ence on Machine Learning (2024) 7

    Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vi- sion mamba: E fficient visual representation learning with bidi- rectional state space model. In: Forty-first International Confer- ence on Machine Learning (2024) 7

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.