pith. machine review for the scientific record. sign in

arxiv: 2604.15174 · v2 · submitted 2026-04-16 · 💻 cs.LG · cs.AI

Recognition: unknown

MambaSL: Exploring Single-Layer Mamba for Time Series Classification

Yoo-Min Jung , Leekyung Kim
Authors on Pith no claims yet

Pith reviewed 2026-05-10 11:24 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords time series classificationstate space modelsMambasingle-layer architectureUEA datasetsreproducible evaluationselective SSM
0
0 comments X

The pith

A minimally redesigned single-layer Mamba model outperforms 20 baselines on all 30 UEA time series classification datasets.

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

The paper shows that a single-layer state space model can serve as a strong backbone for time series classification after targeted changes to its selective state space and projection components. Four hypotheses specific to time series tasks guide these changes, which the authors test by re-running 20 prior methods under one consistent protocol across every UEA dataset. The resulting MambaSL version records higher average accuracy with statistical significance and supplies public checkpoints for every model tested. This matters because it indicates that shallow, efficient sequence models may be sufficient for classification once the architecture is aligned with the problem's demands, while also addressing long-standing issues of incomplete benchmarks and non-reproducible setups.

Core claim

MambaSL applies four TSC-specific hypotheses to minimally redesign the selective SSM block and the projection layers inside a single-layer Mamba. When evaluated against 20 strong baselines on all 30 UEA datasets under a single unified protocol, the model records the highest average accuracy with statistically significant gains and releases public checkpoints for every compared model. Visualizations further illustrate that the adapted single-layer structure functions effectively as a time series classification backbone.

What carries the argument

MambaSL, the single-layer selective state space model whose SSM and projection layers are altered according to four time series classification hypotheses.

If this is right

  • Single-layer sequence models become competitive for time series classification once the selective mechanism and output projections are aligned with task demands.
  • Unified protocols across the full set of UEA datasets become the minimum standard for claiming superiority among time series methods.
  • Public release of trained checkpoints for every baseline allows direct verification and incremental improvement by later researchers.
  • Shallow Mamba variants can serve as efficient starting points for other temporal pattern recognition tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar hypothesis-driven tweaks could be tested on longer or irregularly sampled time series where computational cost grows quickly with depth.
  • The selective state update in Mamba may prove especially useful for capturing class-discriminative temporal motifs without needing stacked layers.
  • The approach invites direct comparison of single-layer Mamba against single-layer transformers on the same reproducible benchmark.

Load-bearing premise

The observed accuracy gains result from the four hypothesis-driven architectural changes rather than from hyperparameter choices or the particular unified protocol applied to the 30 datasets.

What would settle it

An independent run that uses the released public checkpoints but applies a different yet still unified evaluation protocol across the same 30 datasets and finds no statistically significant improvement.

Figures

Figures reproduced from arXiv: 2604.15174 by Leekyung Kim, Yoo-Min Jung.

Figure 1
Figure 1. Figure 1: TI/TV parameterization of (left) ∆ and (right) B in the SSM. TI-∆ fixes the update rate, while TV-∆ adapts it to align sequences with varying speeds. TI-B preserves channel independence, whereas TV-B introduces input-dependent mixing. C follows the same pattern as B at the output stage, highlighting the temporal pacing of ∆ versus the spatial routing of B and C. While the parameters in Mamba share the same… view at source ↗
Figure 2
Figure 2. Figure 2: Overall structure of MambaSL, a single-layer Mamba framework designed for TSC. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of proposed multi-head adaptive pooling with (L, dm, Nh, dy) = (7, 2, 3, 2). H4: Aggregate via adaptive pooling After the Mamba block, per-timestep features are aggregated into a logit vec￾tor l. Conventional pooling methods, such as average or max, treat all steps equally or rely on a single dominant one, thus ig￾noring data-specific temporal importance. This issue is partic￾ularly critical f… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of average classification performances on the UEA benchmark. Non-DL mod [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of the UEA classification results of [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of the UEA classification results us [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of classification accuracy of MambaSL along TI/TV configurations on (left) [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of adaptive pooling on the [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of classification accuracy of MambaSL along TI/TV configurations on the [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: (Continued.) Visualization of classification accuracy of MambaSL along TI/TV configu￾rations on the 30 UEA datasets, in order of maximum accuracy. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: (Continued.) Visualization of classification accuracy of MambaSL along TI/TV configu￾rations on the 30 UEA datasets, in order of maximum accuracy. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
read the original abstract

Despite recent advances in state space models (SSMs) such as Mamba across various sequence domains, research on their standalone capacity for time series classification (TSC) has remained limited. We propose MambaSL, a framework that minimally redesigns the selective SSM and projection layers of a single-layer Mamba, guided by four TSC-specific hypotheses. To address benchmarking limitations -- restricted configurations, partial University of East Anglia (UEA) dataset coverage, and insufficiently reproducible setups -- we re-evaluate 20 strong baselines across all 30 UEA datasets under a unified protocol. As a result, MambaSL achieves state-of-the-art performance with statistically significant average improvements, while ensuring reproducibility via public checkpoints for all evaluated models. Together with visualizations, these results demonstrate the potential of Mamba-based architectures as a TSC backbone.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes MambaSL, a single-layer Mamba architecture for time series classification that applies a minimal redesign to the selective SSM and projection layers guided by four TSC-specific hypotheses. To address prior benchmarking limitations, it re-evaluates 20 baselines across all 30 UEA datasets under a single unified protocol and reports state-of-the-art average performance with statistical significance, supported by public checkpoints for all models.

Significance. If the reported gains can be attributed to the architectural changes rather than protocol differences, the work would establish single-layer Mamba variants as a competitive and efficient TSC backbone, with the unified re-evaluation and public checkpoints providing a valuable reproducibility contribution to the field.

major comments (2)
  1. [Experiments section] The central SOTA claim (abstract and experimental results) attributes average improvements to the four TSC hypotheses' redesign of the selective SSM and projection layers, yet no ablation experiments isolate these changes from the unified protocol (e.g., fixed splits, optimizer, early stopping, or normalization). Without such controls, it is impossible to confirm that gains arise from the architecture rather than the new evaluation setup applied to baselines.
  2. [Methodology section] The method description states that the redesign is 'minimal' and 'guided by four TSC-specific hypotheses,' but provides no equations, pseudocode, or explicit mapping from each hypothesis to the concrete modifications in the SSM or projection layers. This omission is load-bearing for verifying the claim that the changes are both minimal and TSC-specific.
minor comments (2)
  1. [Abstract and Results] The abstract and results mention 'statistically significant average improvements' but do not specify the exact test (e.g., paired t-test, Wilcoxon), correction for multiple comparisons, or per-dataset breakdowns that would allow readers to assess robustness.
  2. [Figures] Figure captions and axis labels in the visualizations could be expanded to explicitly link observed patterns back to the four hypotheses.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the detailed review. We appreciate the feedback on clarifying the contributions of our architectural redesigns and improving the methodological description. We address each major comment below.

read point-by-point responses

  1. Referee: [Experiments section] The central SOTA claim (abstract and experimental results) attributes average improvements to the four TSC hypotheses' redesign of the selective SSM and projection layers, yet no ablation experiments isolate these changes from the unified protocol (e.g., fixed splits, optimizer, early stopping, or normalization). Without such controls, it is impossible to confirm that gains arise from the architecture rather than the new evaluation setup applied to baselines.

    Authors: We agree that additional controls would strengthen the attribution of performance gains to the proposed redesigns. The unified protocol was introduced to enable fair and reproducible comparisons, and all baselines were re-evaluated under identical conditions. However, to directly address this concern, we will include ablation experiments in the revised manuscript. These will evaluate MambaSL variants with individual hypothesis-driven modifications disabled, all under the same unified protocol, to isolate their contributions. revision: yes

  2. Referee: [Methodology section] The method description states that the redesign is 'minimal' and 'guided by four TSC-specific hypotheses,' but provides no equations, pseudocode, or explicit mapping from each hypothesis to the concrete modifications in the SSM or projection layers. This omission is load-bearing for verifying the claim that the changes are both minimal and TSC-specific.

    Authors: We acknowledge that the current manuscript lacks sufficient detail in mapping the hypotheses to specific changes. In the revision, we will expand the Methodology section to include the mathematical formulations of the modified selective SSM and projection layers, pseudocode for the overall architecture, and a clear table or list explicitly linking each of the four TSC-specific hypotheses to the corresponding modifications. This will allow readers to verify the minimal and targeted nature of the redesigns. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical claims rest on external benchmarks

full rationale

The manuscript contains no equations, derivations, or fitted parameters that could reduce to self-definitions or predictions by construction. The central claim is an empirical comparison of MambaSL against 20 baselines on the 30 UEA datasets under a unified protocol, supported by public checkpoints. No self-citation chains, ansatzes, or uniqueness theorems are invoked to justify the architecture or results. The four TSC hypotheses guide a minimal redesign, but this is presented as a design choice rather than a mathematical necessity that collapses into the inputs. This is the expected non-finding for an empirical architecture paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on the assumption that the UEA benchmark is representative and that the four hypotheses capture the key requirements for TSC; no new entities or fitted constants are introduced in the abstract.

axioms (1)
  • domain assumption Mamba selective SSM can be minimally adapted for TSC via four domain hypotheses
    Invoked to justify the redesign choices in the abstract.

pith-pipeline@v0.9.0 · 5436 in / 1158 out tokens · 36896 ms · 2026-05-10T11:24:28.131118+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 10 canonical work pages · 4 internal anchors

  1. [1]

    TimeMachine: A time series is worth 4 mambas for long-term forecasting

    Md Atik Ahamed and Qiang Cheng. TimeMachine: A time series is worth 4 mambas for long-term forecasting. InECAI 2024: 27th European Conference on Artificial Intelligence, volume 392, pp. 1688–1965,

    2024

  2. [2]

    Anthony Bagnall, Hoang Anh Dau, Jason Lines, Michael Flynn, James Large, Aaron Bostrom, Paul Southam, and Eamonn Keogh

    ISSN 2835-8856. Anthony Bagnall, Hoang Anh Dau, Jason Lines, Michael Flynn, James Large, Aaron Bostrom, Paul Southam, and Eamonn Keogh. The UEA multivariate time series classification archive, 2018,

    2018

  3. [3]

    A., Lines, J., Flynn, M., Large, J., Bostrom, A.,

    URLhttps://arxiv.org/abs/1811.00075. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate,

    work page arXiv

  4. [4]

    Neural Machine Translation by Jointly Learning to Align and Translate

    URLhttps://arxiv.org/abs/1409.0473. Ac- cepted at ICLR 2015 as oral presentation. Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    work page internal anchor Pith review arXiv 2015

  5. [5]

    An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling

    URLhttps://arxiv.org/abs/ 1803.01271. Donald J Berndt and James Clifford. Using dynamic time warping to find patterns in time series. In Proceedings of the 3rd International Conference on Knowledge Discovery and Data Mining, pp. 359–370,

    work page internal anchor Pith review arXiv

  6. [6]

    The UCR time series classification archive, July 2015.www.cs.ucr.edu/ ˜eamonn/time_series_data/

    Yanping Chen, Eamonn Keogh, Bing Hu, Nurjahan Begum, Anthony Bagnall, Abdullah Mueen, and Gustavo Batista. The UCR time series classification archive, July 2015.www.cs.ucr.edu/ ˜eamonn/time_series_data/. Tri Dao and Albert Gu. Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality. InProceedings of the 41...

    2015

  7. [7]

    Schmidt, and Geoffrey I

    11 Published as a conference paper at ICLR 2026 Angus Dempster, Daniel F. Schmidt, and Geoffrey I. Webb. Minirocket: A very fast (almost) de- terministic transform for time series classification. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pp. 248–257,

    2026

  8. [8]

    InceptionTime: Finding alexnet for time series classification.Data Mining and Knowledge Discovery, 34(6):1936–1962,

    Hassan Ismail Fawaz, Benjamin Lucas, Germain Forestier, Charlotte Pelletier, Daniel F Schmidt, Jonathan Weber, Geoffrey I Webb, Lhassane Idoumghar, Pierre-Alain Muller, and Franc ¸ois Petit- jean. InceptionTime: Finding alexnet for time series classification.Data Mining and Knowledge Discovery, 34(6):1936–1962,

    1936

  9. [9]

    Residual LSTM: Design of a deep recurrent architecture for distant speech recognition

    Jaeyoung Kim, Mostafa El-Khamy, and Jungwon Lee. Residual LSTM: Design of a deep recurrent architecture for distant speech recognition. InInterspeech 2017, pp. 1591–1595,

    2017

  10. [10]

    VideoMamba: State space model for efficient video understanding

    Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. InComputer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part XXVI, pp. 237–255,

    2024

  11. [11]

    MTS-Mixer: Multivariate time series forecasting via factorized temporal and channel mixing,

    12 Published as a conference paper at ICLR 2026 Zhe Li, Zhongwen Rao, Lujia Pan, and Zenglin Xu. MTS-Mixer: Multivariate time series forecasting via factorized temporal and channel mixing,

    2026

  12. [12]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    URLhttps://arxiv.org/abs/1802.03426. Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language model- ing via gated state spaces. InThe Eleventh International Conference on Learning Representations, 2023a. Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language model- ing via gated state spaces. InThe ...

    work page internal anchor Pith review arXiv

  13. [13]

    k-Shape: Efficient and accurate clustering of time series

    John Paparrizos and Luis Gravano. k-Shape: Efficient and accurate clustering of time series. In Proceedings of the 2015 ACM SIGMOD International Conference on Management of Data, pp. 1855–1870,

    2015

  14. [14]

    MultiRocket: multiple pooling operators and transformations for fast and effective time series classification.Data Mining and Knowledge Discovery, 36:1623–1646,

    13 Published as a conference paper at ICLR 2026 Chang Wei Tan, Angus Dempster, Christoph Bergmeir, and Geoffrey I Webb. MultiRocket: multiple pooling operators and transformations for fast and effective time series classification.Data Mining and Knowledge Discovery, 36:1623–1646,

    2026

  15. [15]

    Deep Time Series Models: A Comprehensive Survey and Benchmark

    Shiyu Wang, Jiawei LI, Xiaoming Shi, Zhou Ye, Baichuan Mo, Wenze Lin, Ju Shengtong, Zhixuan Chu, and Ming Jin. TimeMixer++: A general time series pattern machine for universal predictive analysis. InThe Thirteenth International Conference on Learning Representations, 2025a. Yihe Wang, Nan Huang, Taida Li, Yujun Yan, and Xiang Zhang. Medformer: A multi-gra...

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Etsformer: Exponential smoothing transformers for time-series forecasting.arXiv preprint arXiv:2202.01381, 2022

    URLhttps://arxiv.org/abs/ 2202.01381. Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. TimesNet: Temporal 2d-variation modeling for general time series analysis. InThe Eleventh International Conference on Learning Representations,

    work page arXiv

  17. [17]

    org/abs/2207.01186

    URLhttps://arxiv. org/abs/2207.01186. Yunhao Zhang and Junchi Yan. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. InThe Eleventh International Conference on Learning Representations,

    work page arXiv

  18. [18]

    Most were run on NVIDIA GTX 1080 Ti (11GB), while a few required NVIDIA A100 (40GB) on Google Colab due to memory limits

    14 Published as a conference paper at ICLR 2026 A EXPERIMENTAL SETUP EnvironmentAll experiments were implemented in Python 3.12.8 and PyTorch 2.5.1. Most were run on NVIDIA GTX 1080 Ti (11GB), while a few required NVIDIA A100 (40GB) on Google Colab due to memory limits. Classical baselines used theaeontoolkit (Middlehurst et al., 2024), and all deep model...

    2026

  19. [19]

    AsTSLibhas become a widely used framework, a subset of 10 datasets (EC,FD,HW,HB,JV,PEMS,SRS1,SRS2,SAD, andUW) is commonly used in recent TSC benchmarking practices

    provides 30 multivariate TSC datasets with di- verse sample sizes, input dimensions, lengths, and class counts (Table 4). AsTSLibhas become a widely used framework, a subset of 10 datasets (EC,FD,HW,HB,JV,PEMS,SRS1,SRS2,SAD, andUW) is commonly used in recent TSC benchmarking practices. Table 4: Summary of the 30 UEA datasets used in our experiments Datase...

    1945

  20. [20]

    •Train epochs: 100 •Patience: 10 •Dropout: 0.1 •Seed: 2021 (for DL models) Although this unified setting may deviate from model-specific defaults, extensive model-level grid searches compensated for potential performance degradation (see appendix C.3). B.2 MODEL HYPERPARAMETER SETTINGS FOR GRID SEARCH We primarily considered the hyperparameter settings th...

    2021

  21. [21]

    1”, “13”, “5

    –MTS-Mixer: 256 combinations * e layers: 2 * d model: 128, 256, 512, 1024 * d ff: 0, 2, 4, 8, 16, 32, 64, 128 * fac C: 0 (False) ifd ffis 0, else 1 (True) * down sampling window: 0, 1%, 2%, 3%, 5%, 7.5%, 10%, 12.5% of sequence length * fac T: 0 (False) ifdown sampling windowis 0, else 1 (True) * use norm: 1 (True) 16 Published as a conference paper at ICL...

    2026

  22. [22]

    *” indicates “former

    Also, note that only one hyperparameter combination was tested in our limited resource environment for TSCMamba (Ahamed & Cheng, 2025)–IWpair, since preprocessing took over 2 days and training took over 12 hours per epoch. Table 5: Classification accuracy (%) of TSC models on 30 datasets from the UEA archives. The best and the second-best are highlighted ...

    work page arXiv 2025

  23. [23]

    Model rankings were com- puted within the ablated variants, while additional experimental results are included for reference

    For a clear comparison, we also performed the hyperparameter grid search for each ablated model. Model rankings were com- puted within the ablated variants, while additional experimental results are included for reference. Since removing H2 reduces the number of valid hyperparameter configurations to one-eighth, we extended the search space of Mamba’s hyp...

    work page arXiv 2026

  24. [24]

    report≥ours

    were used. Note that we approximated the reported values using the number of test samples since they were typically rounded to one or two decimal places. The metrics used in Table 7 are as follows: •report≤ours:Count of datasets on which our experimental results of grid search achieved higher or equal accuracy compared to the reported results. A value hig...

    1994

  25. [25]

    Although using three layers yielded the highest average accuracy on the 10 datasets commonly evaluated inTSLib, this observation does not hold when extended to the full set of 30 UEA datasets. When examining the av- erage rank on these 10 datasets, we find that the single-layer configuration consistently outperforms the multi-layer variants, indicating th...

    work page arXiv 2026

  26. [26]

    In contrast to Table 9, where HC2 and Hydra showed the strongest performance based on the best- performing trial, averaging across all trials indicates that MR+Hydra achieves the strongest mean performance and exhibits more consistent results. Nevertheless, MambaSL remains competitive, outperforming all conventional non-DL baselines in direct head-to-head...

    2020

  27. [27]

    Table 11: Summary of ADFTD and FLAAP datasets Dataset Domain Samples Length Variables Classes File Size ADFTD (2023) EEG 69752 256 3 19 2.52GB FLAAP (2022) HAR 13123 100 10 6 60.2MB Preprocessing and experimental settings were aligned with the official Medformer implementation. We first evaluated MambaSL using the 240 hyperparameter configurations employe...

    2023