arXiv CSopen access

MambaSL: Exploring Single-Layer Mamba for Time Series Classification

2026 · arxiv_cs

arXiv CS · Papers · License: Open Access · 2026

machine learning, deep learning, neural networks

Published as a conference paper at ICLR 2026

M AMBA SL: E XPLORING S INGLE -L AYER M AMBA FOR T IME S ERIES C LASSIFICATION Yoo-Min Jung & Leekyung Kim Department of Industrial Engineering, Seoul National University [email protected], [email protected]

arXiv:2604.15174v1 [cs.LG] 16 Apr 2026

A BSTRACT 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.

I NTRODUCTION

State space models (SSMs) have emerged as a strong alternative to Transformers (Vaswani et al., 2017), with Mamba (Gu & Dao, 2024; Dao & Gu, 2024) marking a milestone in sequence domains such as language and video (Lenz et al., 2025; Li et al., 2024). In time series, however, convolutionaland Transformer-based architectures still dominate, with the former excelling in time series classification (TSC) and the latter in time series forecasting (TSF) (Wang et al., 2024b). While early works such as TimeMachine (Ahamed & Cheng, 2024) and S-Mamba (Wang et al., 2025b) showed Mamba’s promise in TSF, its role in TSC remains underexplored. Two gaps motivate our study. First, the standalone capacity of Mamba for TSC has seen little investigation. Wang et al. (2024b) ranked Mamba as the weakest TSC backbone, likely due to a lack of research rather than inherent architectural limitations, as only the vanilla variant was evaluated. To the best of our knowledge, TSCMamba (Ahamed & Cheng, 2025) is the only subsequent benchmarking attempt for TSC. However, its integration of feature engineering techniques such as ROCKET (Dempster et al., 2020) and CWT (Mallat, 1999) obscures Mamba’s intrinsic contribution. As a second gap, current TSC benchmarking suffers from three critical issues related to coverage, fairness, and reproducibility. First, evaluations are often restricted to only a fraction of the University of East Anglia (UEA) archive (Bagnall et al., 2018), leaving out challenging datasets with long sequences or high input dimensionality. Second, non-TSC models are frequently adopted without proper re-tuning, which risks underestimating their capacity; for example, TSF models like DLinear (Zeng et al., 2023) and PatchTST (Nie et al., 2023) have been reported in TSC contexts using default settings (Wu et al., 2023; Luo & Wang, 2024). Finally, reproducibility remains a concern— e.g., re-evaluations of TS2Vec (Yue et al., 2022) and GPT4TS (Zhou et al., 2023) by Eldele et al. (2024) revealed average accuracy drops exceeding 9%p compared to their original reports. These limitations undermine the reliability of comparative conclusions in the literature. Motivated by these gaps, we revisit Mamba as a TSC backbone from two perspectives. (i) Architecture: We propose four TSC-specific hypotheses (H1–H4), which guide the redesign of selective SSM components and the input and output projection layers. (ii) Evaluation protocol: We establish a consistent benchmarking setup and re-evaluate strong TSC baselines across all 30 multivariate UEA datasets with extensive hyperparameter sweeps. Under the protocol, our single-layer Mamba TSC framework, MambaSL, achieves state-of-the-art performance. Our key contributions are: 1

Published as a conference paper at ICLR 2026

• TSC-specific hypotheses and architectural refinements. We propose the following four hypotheses (H1–H4) that explain why naively applying Mamba within the existing TSC pipeline does not fully realize its potential, and validate them through corresponding design adjustments: (H1) Scale input projection. Since Mamba’s output is modulated by a gating unit (Hua et al., 2022; Mehta et al., 2023a), insufficient input context can bottleneck performance, motivating a larger input projection receptive field for densely sampled time series. (H2) Modularize time (in)variance. As time series often exhibit near linear time-invariant behavior, we decouple time variance of Mamba as a hyperparameter. Simpler configurations often perform better, contradicting the ablation results from Gu & Dao (2024). (H3) Remove skip connection. In shallow networks, skip connections yield minimal performance gains (He et al., 2016; Ismail Fawaz et al., 2020). Given Mamba’s strong long-range memory, we remove skip connections and construct logits solely from hidden states. (H4) Aggregate via adaptive pooling. Time series classification spans both global and eventdriven patterns, which conventional pooling cannot accommodate. We therefore propose a multi-head adaptive pooling that weights temporal features in a dataset-specific manner. • Comprehensive and reproducible UEA benchmarking. We evaluate 20 models on all 30 UEA datasets, covering various sequence lengths (8–17,984), input dimensions (2–1,345), and sample sizes (12–25,000). For each model, we explore approximately 200 hyperparameter combinations to select the optimal configuration. Notably, TSF models previously tested on the UEA datasets showed an average accuracy improvement of 3.04%p with hyperparameter tuning alone. • Empirical validation of a single-layer Mamba. Our MambaSL framework achieves the stateof-the-art performance across the UEA benchmark, outperforming the second-best method by 1.41%p. We demonstrate the inherent potential of Mamba by validating our hypotheses through ablation studies and highlighting backbone- and dataset-specific features through visualizations. The remainder of the paper is organized as follows. Section 2 reviews related work for TSC and Mamba, section 3 details our hypotheses and architectural refinements, section 4 describes our experimental setup and implementation details, section 5 presents results and analysis, and section 6 concludes the paper.

R ELATED WORK

Recent advances in deep learning (DL) for TSC can be grouped into three streams. (i) TSC-specific models such as TS2Vec tailor encoders and readouts directly for classification (Yue et al., 2022; Eldele et al., 2024; Wen et al., 2025). (ii) Foundation models aim to adapt a single model across multiple time-series tasks, including TSC (Wu et al., 2023; Zhou et al., 2023; Luo & Wang, 2024; Wang et al., 2025a). (iii) TSF-origin models are frequently repurposed as baselines for TSC-specific and foundation models (Zeng et al., 2023; Zhang & Yan, 2023; Nie et al., 2023). Although evaluations typically rely on the UEA benchmark, inconsistent use of dataset subsets and single-setting evaluations for TSF-origin models have led to underestimated baselines. This highlights the need for standardized and consistent benchmarking in TSC. Within SSMs, Mamba has been explored primarily for TSF. Variants such as TimeMachine (Ahamed & Cheng, 2024) and S-Mamba (Wang et al., 2025b) typically construct temporally mixed embeddings and then apply bidirectional Mamba along the channel axis to mitigate scan-order sensitivity. By contrast, adaptations to TSC in general are comparatively scarce, with TSCMamba (Ahamed & Cheng, 2025) being the only variant. Across TSF and TSC alike, Mamba variants tend to update the SSM state along non-temporal axes (channel or frequency) rather than the original time axis.

M ETHODS

3.1

P ROBLEM D EFINITION

Let X = x1:L = [x1 ; . . . ; xL ] be a multivariate time series of length L, where each xt at time t is a dx -D vector. TSC infers the label y ∈ {1, . . . , dy } from X, where dy is the number of classes. Many TSC architectures consist of three fundamental modules: an input projection ΦI that maps the (1) (d ) input subsequence xt−k+1:t (receptive field of size k) into a dm -D vector x̃t = [ x̃t , . . . , x̃t m ], 2

Published as a conference paper at ICLR 2026

a feature extractor ΦFE that produces per-timestep feature vectors f1:L from x̃1:L (typically of the same size), and an output projection ΦCLF that aggregates f1:L into logits l ∈ Rdy . Formally, x̃t = ΦI xt−k+1:t , (1) f1:L = ΦFE x̃1:L , (2) l = ΦCLF f1:L , (3) ŷ = arg max softmax(l)i .

(4)

i∈{1,...,dy }

The aggregation function of ΦCLF can be instantiated as a fully connected layer, global average/max pooling, or a last-timestep readout. Whereas many recent studies primarily emphasize improving ΦFE , we examine and improve the entire TSC pipeline for Mamba in section 3.3. 3.2

T IME - VARIANT SSM IN M AMBA

At a high level, Mamba can be viewed as a selective SSM core augmented with a lightweight gated linear block (Hua et al., 2022; Mehta et al., 2023a). We next review the selective SSM, clarifying its components and establishing the notation used in our proposed method (section 3.3.2). 3.2.1

L INEAR TIME - INVARIANT SSM

An SSM maps a signal u(t) ∈ R to a ds -D state vector s(t) and projects it onto an output y(t) ∈ R: ṡ(t) = As(t) + Bu(t), ds ×ds

y(t) = Cs(t) + Du(t),

ds ×1

(5)

1×ds

with four parameters, A ∈ R ,B ∈ R ,C ∈ R , and D ∈ R. This yields a linear time-invariant (LTI) system when the parameters are fixed over time. With a step size ∆ > 0, the system is discretized as st = Āst−1 + B̄ut ,

yt = Cst + Dut ,

(6)

where Ā = FA (∆, A) and B̄ = FB (∆, A, B) are discretization rules. Different discretization rules exist; the zero-order hold (ZOH) method is discussed in section 3.2.2. While these forms are standard, we retain them here as the basis for our multivariate extension. In SSM-based DL models, this extension is typically implemented by sharing A across channels and broadcasting the other parameters (Gu et al., 2021; 2022; Gu & Dao, 2024). For instance, when (j) using the SSM as the feature extractor ΦFE in equation 2, each input channel x̃t is mapped to a ds -D state via channel-specific B (j) (and step size ∆(j) ) and shared A. Thus, at time t, the hidden (j) (j) (j) (1) (d ) state St = [ st ; . . . ; st m ] ∈ Rds ×dm where st = FA (∆(j) , A)st−1 + FB (∆(j) , A, B (j) )x̃t . (j) (j) The feature vector ft is also generated by channel-specific C and D . That is, in DL practice, the multivariate SSM is fundamentally a channel-independent LTI system. To our knowledge, however, existing works rarely formalize how ∆, B, and C operate across channels. We show explicitly that ∆ primarily acts along the temporal axis on each channel, whereas B and C primarily serve as channel-wise projections under a shared A in Mamba’s selective SSM. This clarification, often left implicit in prior literature, sets the stage for the following section. 3.2.2

S ELECTIVE SSM

Mamba (Gu & Dao, 2024) proposed a selective SSM that allows ∆, B, and C to selectively propagate or forget information from given input. Given the projected input x̃t ∈ Rdm , three learnable maps, ϕ∆ , ϕB , and ϕC , produce the input-dependent parameters ∆t , Bt , and Ct , respectively: (1) (d ) m , ∆t = ϕ∆ (x̃t ) = ζ Linearbias = (∆t , . . . , ∆t m ) ∈ Rd>0 dm Lineardr (x̃t ) ds ×1 (7) Bt = ϕB (x̃t ) = Lineards (x̃t ) ∈R , Ct = ϕC (x̃t ) = Lineards (x̃t )⊤

∈ R1×ds .

d Here, ζ is the softplus activation, Linearbias d (x) and Lineard (x) denote linear projections into R d×1 (identified with R ) with and without bias, and dr is the rank of low-rank projection for ∆t .

Published as a conference paper at ICLR 2026

The selective SSM is then formulated as follows:1 (j) (j) (j) (j) (j) (j) (j) (j) st = Āt st−1 + B̄t x̃t , ft = Ct st + D (j) x̃t ,

j = 1, . . . , dm ,

(8)

(j) (j) where Āt and B̄t are discretized by the ZOH rule (Mehta et al., 2023b; Gu & Dao, 2024): (j) (j) (j) Āt = FA (∆t , A) = exp(∆t A), (j) (j) (j) (j) (j) B̄t = FB (∆t , A, Bt ) = (∆t A)−1 (exp(∆t A) − I)∆t Bt .

(9)

This transformation changes the SSM from an LTI to a time-variant (TV) system, enabling contextaware reasoning that selectively stores or extracts key information. The explicit formulation also (j) (j) highlights that A, Bt , and Ct are shared across channels, while Āt and B̄t differ due to the (j) channel-wise ∆t . O N TIME VARIANCE IN ∆, B, AND C

3.2.3

TI-Δ

Δ 2 Δ 3 Δ 4 𝑡

TV-Δ

Δ!

TI-B

Δ" 2 Δ$ 3 Δ# 4 𝑡

𝐵! 𝐵 "

TV-B

𝑊&

𝐵%

𝐵! 𝑥"#! 𝐵" 𝑥"#" 0

Δ 2 Δ 3 Δ 4 𝑡

Δ! 1 Δ" 2 Δ$

𝑥"#! 𝑥"#"

Δ# 4 𝑡

𝐵# 𝑥"#! 𝐵# 𝑥"#"

𝑥#%

𝑥#%!

𝑥#%"

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 mechanism to yield a TV system by default, as defined in equation 7, their main roles differ. We denote ∆, B, and C in the LTI SSM using TInotation (section 3.2.1), and ∆t , Bt , and Ct in the selective SSM using TV- notation (section 3.2.2). ∆: Temporal Update Rate. ∆ controls the timescale of state updates. TI-∆ keeps a constant rate, whereas TV-∆ adapts to context: larger ∆ accelerates updates, and smaller ∆ prolongs memory. This resembles dynamic time warping (DTW), aligning sequences with varying local speeds (see the left side of Figure 1). B: Input-to-State Routing. B determines how each input channel drives the latent state. TI-B enforces channel-independent routing, whereas TV-B enables context-dependent mixing of input channels before entering the state space (see the right side of Figure 1). C: State-to-Output Readout. C maps states to output features. TI-C applies a fixed readout, preserving channel independence. TV-C introduces adaptive mixing at the output stage, potentially capturing richer cross-channel interactions. Interplay. TV-∆ governs temporal dynamics, whereas TV-B and TV-C govern spatial mixing. The extreme cases are: • TI-∆, TI-B, TI-C: LTI system with channel-independent I/O; • TV-∆, TV-B, TV-C: fully TV system with channel-mixed I/O. Thus, temporal and spatial variability can be decoupled and selectively controlled, motivating our second hypothesis (H2) in section 3.3.2. Note that, while these roles are primary, the ZOH rule couples them to some extent: ∆ can indirectly modulate the influence of B and C, and vice versa. 3.3

S INGLE - LAYER M AMBA FOR TSC

We refine the vanilla Mamba and projection layers through four hypotheses (H1–H4) motivated by their limitations in the standard TSC pipeline. The full framework is shown in Figure 2. 1 For simplicity, we treat the selective SSM without shift as the feature extractor itself, using x̃t in place of ut and ft in place of yt from equation 6. In practice, ∆, B, and C are generated from an expanded input after a linear projection and a local convolution, but we omit this detail to keep the notation concise.

Published as a conference paper at ICLR 2026

Output ∑

SSM Output Projection A

per-time weight

per-time logit

Softmax

Linear

σ Conv1D

Modular Selective SSM

Linear

Dropout

Adaptive Max Pooling

Linear

Linear σ LayerNorm

Dropout Positional Encoding

MambaBlock ×1

Input Projection

: SiLU : Depends on time variability hyperparameters (𝜃! , 𝜃" , 𝜃# )

Conv1D

Input

: Learnable parameter

Figure 2: Overall structure of MambaSL, a single-layer Mamba framework designed for TSC. 3.3.1

I NPUT PROJECTION

H1: Scale input projection In recent time series models, ΦI is implemented as a 1-D convolution with a fixed kernel size k = 3 (Wu et al., 2023; Zhou et al., 2023), unless alternatives such as patching or frequency-domain transforms are employed (Nie et al., 2023; Wang et al., 2025a). Since Mamba’s gating mechanism modulates the SSM output based on this projection, we hypothesize that longer sequences require proportionally larger receptive fields. We therefore define k = max(kmin , ⌊λL⌋),

(10)

with kmin = 3 (minimum k), λ = 0.02 (sequence ratio), and stride = 1 to isolate kernel-size effects. 3.3.2

F EATURE EXTRACTOR

We adopt the vanilla Mamba as our feature extractor, introducing minimal changes to its SSM core under two hypotheses: (H2) modularizing time (in)variance and (H3) removing the skip connection. H2: Modularize time (in)variance Building on section 3.2.3, we hypothesize that the optimal TI/TV configuration of ∆, B, and C is dataset-dependent. Evidence from TSF shows that channel independence may outperform mixing (Zeng et al., 2023; Nie et al., 2023), and shape-based distance—akin to TI-∆—can rival DTW in TSC (Paparrizos & Gravano, 2015). As aggregation with channel mixing is inevitable at the output, we expect C to have a comparatively smaller effect, while ∆ and B are likely to exert a larger dataset-specific influence. To systematically examine this, we introduce binary switches θ∆ , θB , θC ∈ {0, 1} that determine whether each parameter follows a TI or TV form. For channel j ∈ {1, . . . , dm } and time t, the effective parameters are (j) ⋆

∆t

(j)

= (1 − θ∆ ) ∆(j) + θ∆ ∆t

(j) ⋆ Bt = (1 − θB ) B (j) + θB Bt (j) ⋆ Ct = (1 − θC ) C (j) + θC Ct

= (1 − θ∆ ) ∆(j) + θ∆ ϕ∆ (x̃t )(j) ∈ R>0 , = (1 − θB ) B (j) + θB ϕB (x̃t )

∈ Rds ×1 ,

= (1 − θC ) C (j) + θC ϕC (x̃t )

∈ R1×ds ,

(11)

where all ϕ are defined in equation 7. This modularization yields 23 = 8 configurations, ranging from a fully LTI system (all θ = 0) to the selective SSM (all θ = 1). Substituting equation 11 into equations 8 and 9 produces our modularized selective SSM, illustrated in the upper left of Figure 2. 5

Published as a conference paper at ICLR 2026

H3: Remove skip connection Skip (residual) connections improve optimization in deep networks, yet their effect diminishes in shallow networks (He et al., 2016; Kim et al., 2017). InceptionTime (Ismail Fawaz et al., 2020) further shows that the skip connection makes little difference across 85 UCR datasets (Chen et al., 2015), indicating that it is not always essential for TSC. In the single-layer setting, such shortcuts may bypass the SSM and hinder Mamba’s representation learning. We therefore hypothesize that removing the skip connection forces the model to rely solely (j) on the SSM’s state evolution. This is implemented by omitting the D (j) x̃t term in equation 8 2 : (j)

(j)

= Ct st

(no skip term) .

(12)

(j) This modification positions learning the state vector st as the core of TSC.

3.3.3

O UTPUT PROJECTION

H4: Aggregate via adaptive pooling After the Mamba block, per-timestep features are aggregated into a logit vector l. Conventional pooling methods, such as average or max, treat all steps equally or rely on a single dominant one, thus ignoring data-specific temporal importance. This issue is particularly critical for recurrent models, where the predicted label may shift over time (e.g., a Handwriting sequence labeled g may initially resemble class a before later aligning with g). We propose a multi-head adaptive pooling with learnable gates (Figure 3). Specifically, Nh independent gating heads each produce a scalar score gt,h for time step t: gt,h = wh⊤ ft + bh ,

h = 1, . . . , Nh . (13) For each t, the maximum gate value across heads is selected and normalized via softmax to obtain per-timestep weight αt , which is then applied to the per-timestep logit vector lt : gt = max gt,h , h

exp(gt ) αt = PL , i=1 exp(gi )

𝒇!:# 𝑔!:#,!:%! Linear

Linear

𝒍!:# 𝑔!:#

max

𝛼!:#

softmax

per-time logit

0% 0% 3% 6% 25% 44% 9% per-time weight

𝒍

Figure 3: Illustration of proposed multi-head adaptive pooling with (L, dm , Nh , dy ) = (7, 2, 3, 2). l=

L X

αt lt .

(14)

t=1

This formulation generalizes conventional pooling: uniform αt recovers averaging, while a sharply peaked αt approximates max pooling. Compared to attention pooling (Bahdanau et al., 2016), our design is lightweight yet expressive: multi-head gating explores diverse patterns, while adaptive max pooling selects the most confident signals, enabling robust dataset-specific aggregation.

E XPERIMENT

We conduct experiments on the full UEA benchmark (Bagnall et al., 2018), covering diverse sequence lengths, input dimensions, and sample sizes. For conciseness, we will hereafter refer to each dataset by its code (e.g., EC for EthanolConcentration; see Table 4). To ensure fairness, we allocate an identical search budget per model–dataset under a unified protocol, selecting the best configuration for each dataset. Detailed descriptions of the environment, datasets, and metrics are provided in appendix A, and hyperparameter settings are provided in appendix B. Baselines We include (1) non-DL methods: DTW-based nearest neighbor (DTWD ; Berndt & Clifford, 1994), ROCKET (Dempster et al., 2020), HIVE-COTE 2.0 (HC2; Middlehurst et al., 2021), Hydra (Dempster et al., 2023), MultiRocket (MR; Tan et al., 2022)+Hydra; (2) MLP-based models: DLinear (Zeng et al., 2023), LightTS (Zhang et al., 2022), MTS-Mixer (Li et al., 2023); (3) CNNbased models: TimesNet (Wu et al., 2023), ModernTCN (Luo & Wang, 2024), TSLANet (Eldele et al., 2024), TimeMixer++ (Wang et al., 2025a); (4) Transformer-based models: FEDformer (Zhou et al., 2022), ETSformer (Woo et al., 2022), Crossformer (Zhang & Yan, 2023), PatchTST (Nie et al., 2023), GPT4TS (Zhou et al., 2023), iTransformer (Liu et al., 2024); (5) shape-based model: InterpGN (Wen et al., 2025); and (6) Mamba-based model: TSCMamba (Ahamed & Cheng, 2025). 2 While Mamba and related literature typically derive the SSM equation without the skip connection D, their official implementations default to enabling it (state-spaces, 2025). We changed this to be tunable.

Published as a conference paper at ICLR 2026

R ESULTS AND DISCUSSION

5.1

C LASSIFICATION RESULTS ON THE UEA BENCHMARK CNN-based 10 dataset all dataset

Transformer-based 10 dataset all dataset

Shape-based 10 dataset all dataset

Mamba-based 10 dataset all dataset 79.82

78.40

75.93

75.48

77.70

74.12

74.80

76.06

74.30

72.89

74.73

75.32

77.43

74.27

75.72

71.54

74.13

10.5 5.3

5.5

7.1

7.9

8.0

8.7

9.5

10.3

11.8

12.6

13.1

8.0

8.5

10.3

10.6

11.8

12.3

6.2

7.0

ET _D CK RO

77.28

7.8

8.2

9.1

7.4

8.7

10.6

11.8

13.9

15.4

13.1

W DT

74.95

73.43

76.26

77.44

76.94

69.00

69.58

16.8

9.4

76.92

75.17

75.31

68.25

16.0

8.8

10.6

8.9

9.6

10.3

11.2

13.2

72.76

71.00

16.0

16.6

76.32

68.02

77.12

77.60

73.31

74.37

78.35

76.87

78.16

78.83

average rank

MLP-based 10 dataset all dataset

70.35

average accuracy (%)

Non-DL 10 dataset all dataset

t t r r r r r r a a a T S N N SL TS ea dr dr tT Ne er++ rme ixe esNe nTC TS me me me pG mb ba T4 Hy +Hy DLin Ligh S-M LA o r x for for nter CMa for tch GP am Df TS de Mi ss ns I Pa TS Tim E M o e S o a E MT MR F r r T M C iT Tim

Figure 4: Comparison of average classification performances on the UEA benchmark. Non-DL models include DTW-D, ROCKET, HC2, Hydra, and MR+Hydra; DL models are grouped by their backbone structures. As shown in Figure 4, our proposed model—MambaSL—established a new state of the art by achieving the best average accuracy and rank across both the widely used 10-dataset subset and the full set of 30 UEA datasets. In particular, MambaSL surpassed the second-best model, TSCMamba, by an average margin of 1.41%p. Wilcoxon signed-rank tests (Wilcoxon, 1945) across the 30 datasets confirmed statistically significant gains (p < 0.05) over all models except HC2 (p = 0.56), despite HC2 ranking only 8th overall (see appendix C.1 for full results and statistics). As detailed in appendix C.1, MambaSL tends to perform strongly on datasets where multi-scale convolutions or frequency-domain features are typically dominant. For example, AWR, CR, EW, EP, HW, PS, RS, UW, and SWJ generally favor non-DL, TSLANet, InterpGN, and TSCMamba, yet MambaSL remains competitive. In particular, for CR and HW, MambaSL is the only DL model that surpasses DTWD . At the same time, MambaSL shows competitive results on datasets such as FD, HMD, NATO, and SRS1, where MLP- or Transformer-based backbones have relative advantages. These results, achieved without additional domain transforms or complex mechanisms, demonstrate both the inherent capability of Mamba and the effectiveness of our architectural refinements. Visualization of TSC models using UMAP (n_neighbors=12, metric=correlation) iTransformer

ModernTCN TimeMixer++MTS-Mixer Crossformer LightTS GPT4TS PatchTST DLinear TimesNet FEDformer ETSformer TSLANet MambaSL TSCMamba

component 2

To further contextualize the results, Figure 5 visualizes each model using a 2-D embedding of its 30 accuracy values via uniform manifold approximation and projection (UMAP) (McInnes et al., 2020). DL and nonDL methods are clearly clustered in separate regions, while InterpGN and MambaSL appear in between. The placement of InterpGN is reasonable since it ensembles shape-based and DL models. This suggests that MambaSL shares the strengths of both DL and non-DL approaches.

InterpGN

avg. acc

DTW_D

Hydra HC2 ROCKET MR+Hydra 8

8.5

9.5

component 1

10.5

11.5

Figure 5: Visualization of the UEA classification results of each TSC model using the UMAP algorithm.

Published as a conference paper at ICLR 2026

Visualization of UEA datasets using UMAP (n_neighbors=9, metric=correlation) 7.5

avg. acc

EW LSST RS HW

7 6.5

component 2

To complement the model-level analysis in Figure 5, we visualize the UEA datasets via UMAP in Figure 6. Here, each point represents a dataset embedded by its 21 accuracy values. Compared to the full results in appendix C.1, this view clearly shows the pattern of non-DL methods; datasets on which all nonDL baselines perform poorly form a tight cluster in the lower right (FD, JV, SRS2, AF, FM, HMD, MI, SWJ), whereas the upper-left cluster tends to outperform under the non-DLs.

LIB

5.5

AWR CT SRS1

SAD DDG

PEMS

NATO

4.5 4 −1

component 1

HB JV

FM MI

FD SWJ

AF HMD SRS2

30 20

Figure 6: Visualization of the UEA classification results using the UMAP algorithm along dataset axis.

Finally, reproducibility remains a challenge in DL models due to randomness in initialization and optimization. We therefore compared our results with prior reports (appendix C.3). Our evaluation yielded higher performance than previously reported results, except for ModernTCN and TimeMixer++. However, their average ranks decline when the reported results are applied, confirming the fairness and rigor of our protocol across all baselines. For ModernTCN, although the average accuracy was lower than expected, we achieved higher accuracy with fewer hyperparameter trials compared to the revisited study (Akacik & Hoogendoorn, 2025). Notably, TSF-origin models (DLinear, LightTS, MTS-Mixer, FEDformer, ETSformer, Crossformer, PatchTST, and iTransformer) improved by over 3%p, suggesting that their methodologies can be reconsidered for TSC. 5.2

M ODEL ABLATION

Table 1: Model ablation for four hypotheses (H1–H4). The best and the second-best are highlighted in bold and underline, respectively.

Model

(H1) kernel size

(H2) (H3) (H4) modular SSM skip connection aggregation

✓: k = 0.02L ✓: modular TV/TI ✗: k = 3 ✗: TV only

✓: not use D ✗: use D

✓: adaptive pool ✗: the others

avg. acc (10)

avg. avg. # win # lose rank acc rank /draw /draw test

MambaSL w/o H1 w/o H2 w/o H3 w/o H4

✓ ✗ ✓ ✓ ✓ ✓ ✓ ✓

✓ ✓ ✗ ✓ ✓ ✓ ✓ ✓

✓ ✓ ✓ ✗ ✓ ✓ ✓ ✓

✓ 78.779 79.802 2.43 ✓ 77.413 80.215 2.53 ✓ 75.659 77.075 5.93 ✓ 77.715 79.163 3.50 ✗ (fully conn) 76.438 77.722 4.87 ✗ (avg pool) 76.900 78.610 3.60 ✗ (max pool) 76.448 78.535 3.90 ✗ (last step) 75.856 76.848 4.87

25 30 22 24 22 22 25

22 5 14 10 15 12 9

only H2 Mamba

✗ ✗

✓ ✗

✗ ✗

✗ (avg pool) 77.092 77.935 ✗ (fully conn) 73.286 74.243

20 26

13 0.011 5 0.000

0.217 0.000 0.071 0.003 0.044 0.004 0.001

Table 1 summarizes the average classification accuracy of ablated models for each hypothesis (H1– H4), providing a systematic exploration of MambaSL’s design space. The full results are included in appendix C.2. MambaSL remained competitive with state-of-the-art baselines, even when excluding H1, H3, or H4 individually. Also, none of the three ablations resulted in a statistically significant gains in the rank test. Motivated by this, we additionally ran an ablation that only applies H2; however, this variant failed to surpass the second-best baseline and showed statistically inferior performance (p = 0.011), suggesting that all our four hypotheses contribute positively to performance. Disabling H1 (k = 3) yielded the highest average accuracy, surpassing 80%. Its overall ranking, however, remained second, as the gain was largely driven by marginal accuracy changes in the AF and SWJ datasets, which both have only 15 test samples. Additionally, on three datasets—AF, ER, and PEMS—where MambaSL was outside the top 10 (appendix C.1), a fully connected layer outperformed other classifiers. This suggests that for fixed-length datasets, fully connected readouts can be competitive as long as overfitting is controlled. Nevertheless, across the full benchmark, our multi-head adaptive pooling consistently performs better with greater generalizability. 8

Published as a conference paper at ICLR 2026

5.3

F URTHER ANALYSIS

Time variance ablation Table 2 shows Table 2: Ablation on time variance of ∆, B, and C. the classification accuracy of MambaSL The best and the second-best are highlighted in bold under eight configurations of time variance and underline, respectively. hyperparameters, θ∆ , θB , and θC . TV-∆ TV-B TV-C avg. acc (10) avg. acc avg. rank ✗ ✗ ✗ ✗ ✓ ✓ ✓ ✓

average accuracy

No single combination is overwhelmingly superior, yet models with the LTI setting (all marked ✗) tend to outperform the fully TV setting (all marked ✓) of selective SSM. This contrasts with the findings of Gu & Dao (2024) in their ablation study, where the fully TV setting was superior for language modeling. 0.4 0.35 0.3

✗ ✗ ✓ ✓ ✗ ✗ ✓ ✓

✗ ✓ ✗ ✓ ✗ ✓ ✗ ✓

77.280 76.938 76.451 76.422 76.253 76.299 76.309 75.938

77.400 77.102 77.042 77.767 77.707 77.148 77.085 77.055

3.87 3.93 3.80 3.60 3.63 3.70 3.80 4.13

0.6

0.55

0.5

TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✓ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓

TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓ configuration

configuration

Figure 7: Visualization of classification accuracy of MambaSL along TI/TV configurations on (left) EC and (right) HW datasets, in order of maximum accuracy. In some datasets, the trend is particularly clear. As shown on the left side of Figure 7, in EC, accuracy increases as more parameters are set to be time-invariant, with ∆ having the largest impact. This indicates that dynamic warping of the sampling intervals can distort the chemical reaction signal of ethanol, which can impair classification performance. Furthermore, similar to the right of Figure 7, using TI-B consistently yields better results than using TV-B across many datasets (HW, HB, CT, DDG, UW, and RS), while the opposite trend is also observed (PEMS, CR, PD, and PS). By contrast, no dataset exhibits a consistent preference with respect to C. These observations support our hypothesis that while the time variance of ∆ and B can be dataset-specific, that of C has limited impact, as output-channel mixing naturally occurs in the classifier. The visualizations for all 30 datasets are provided in appendix C.4. Class label: Badminton_Smash logit value

logit value

Class label: Badminton_Clear 10 0

−10

class

Badminton_Smash Badminton_Clear

−10 0

time step

αt

0.12

Squash_ForehandBoast

0.08

Squash_BackhandBoast

0.04

Figure 8: Visualization of adaptive pooling on the RS dataset for two samples. Line plots show pertimestep logits lt , and heatmaps illustrate the corresponding per-timestep weights at . Analysis of pooling behavior To further examine the role of adaptive pooling, we analyzed the RS dataset, which contains relatively few classes and yields strong performance under max pooling. The left plot in Figure 8 depicts a sample where the predicted label differs between a simple average of per-timestep logits and the weighted average using at . As shown, large logit values do not necessarily correspond to large weights, and regions with high weights are often localized. As in the right plot of Figure 8, there are also samples with nearly uniform weights, effectively reducing adaptive pooling to average pooling. This adaptive weighting enables MambaSL to outperform average pooling, achieving the second-highest accuracy on the RS dataset (appendices C.1 and C.2). Effect of model depth Although our main architecture uses a single Mamba block as shown in Figure 2, we also ablated the number of stacked blocks to examine whether a deeper MambaSL 9

Published as a conference paper at ICLR 2026

further improves performance. Table 3 compares variants with 1, 2, and 3 Mamba blocks under the same training protocol (full results are provided in appendix C.5). We observed that a single layer achieved the best average accuracy and the best average rank across the 30 UEA datasets. While performance gradually decreases as the number of layers increases, the differences are small, and all three variants perform similarly overall. Given these results, we conclude that a carefully configured single-layer Mamba is sufficient as a backbone for TSC in this regime, and we adopt the 1-layer MambaSL as our default architecture for both simplicity and efficiency. Table 3: Classification accuracy (%) of MambaSL according to its model depth. The best and the second-best are highlighted in bold and underline, respectively. depth

avg. acc (10)

avg. acc

avg. rank (10)

avg. rank

# of win/draw

# of lose/draw

rank test

1 2 3

78.827 78.853 78.918

79.819 79.316 79.279

1.30 1.90 2.10

1.33 1.80 2.07

26 24

15 12

0.035 0.004

5.4

A DDITIONAL EXPERIMENTS

UEA benchmark classification considering data leakage issue To account for the data-leakage issue reported for TSLib in TSC settings (Akacik & Hoogendoorn, 2025; Talukder et al., 2024), we additionally trained MambaSL under the InceptionTime protocol (Ismail Fawaz et al., 2020), where the checkpoint with the lowest training loss is used for evaluation. For comparison, we selected the top three high-performing models from each source—HC2 (Middlehurst et al., 2021), BORF (Spinnato et al., 2024), and our own experiments—to form a comparable set of classical baselines. These results, along with the corresponding evaluation tables, are provided in appendix C.6. While MambaSL shows lower accuracy under the stricter InceptionTime protocol than under the default TSLib setting, it remains competitive relative to the strongest non-DL baselines reported in previous studies and in our re-evaluation. Classification on recent domain-specific dataset Although the UEA archive (Bagnall et al., 2018) covers a broad range of TSC datasets, it does not include more recently released real-world benchmarks. Therefore, we additionally evaluated MambaSL on ADFTD (Miltiadous et al., 2023) and FLAAP (Kumar & Suresh, 2022), two of the most recent datasets benchmarked by Medformer (Wang et al., 2024a) in the medical and human activity recognition (HAR) domains, respectively. Classification results compared with strong baselines in Medformer are provided in appendix C.7. Across both datasets, MambaSL achieves competitive performance without any domainspecific modifications, demonstrating its generalizability to recent TSC tasks.

C ONCLUSION

In this study, we introduce four hypotheses as to why Mamba has remained underexplored in TSC despite its success in other sequence domains. With minimal architectural modifications, our singlelayer Mamba framework, MambaSL, achieved state-of-the-art performance across the entire UEA benchmark. By re-examining 20 prior baselines, we further highlighted systematic differences across backbone structures and demonstrated that Mamba can serve as a strong multivariate TSC backbone. Nevertheless, our work has two main limitations. First, while our model is robust across the UEA benchmark, it rarely achieves overwhelming superiority on individual datasets, implying that practical use may require domain-specific adaptations. For instance, prior knowledge of input embedding may constrain kernel size, making receptive-field scaling with sequence length suboptimal in certain domains. Second, our study is limited to the first version of Mamba (Gu & Dao, 2024). While Mamba-2 (Dao & Gu, 2024) introduces an efficient parallel Mamba block, applying time-variance modularization requires low-level optimizations beyond the scope of this study. Overall, our findings demonstrate that a carefully configured single-layer Mamba can serve as a competitive backbone for TSC. Beyond TSC, future work may further improve Mamba by leveraging receptive-field scaling and time-variance control, along with shallow, task-aligned architectural choices and adaptive aggregation strategies. 10

Published as a conference paper at ICLR 2026

R EPRODUCIBILITY STATEMENT Source code and scripts are publicly available at GitHub repository , and datasets, along with the best-performing model checkpoints, have been uploaded to Google Drive. • The source code offers a complete implementation of Mamba and all other baselines, with the exception of dataset and checkpoint downloads. The repository includes grid search script generation, local training execution, pretrained model evaluation, and result visualization. • Full training logs, corresponding to the grid search results reported in Figure 4 and Table 1, are also provided in the repository. We are confident that the experimental setup and hyperparameter settings mentioned in appendices A and B are identical to the scripts and logs inside the repository. • The dataset folder also contains preprocessed CWT and ROCKET features for TSCMamba, compatible with the released checkpoints. • Detailed reproduction instructions are available in the README.md file. T HE USE OF LARGE LANGUAGE MODELS The authors acknowledge the use of ChatGPT-5 and Google Translate only during the writing process to ensure rigorous and concise academic English writing.

R EFERENCES Md Atik Ahamed and Qiang Cheng. TimeMachine: A time series is worth 4 mambas for long-term forecasting. In ECAI 2024: 27th European Conference on Artificial Intelligence, volume 392, pp. 1688–1965, 2024. Md Atik Ahamed and Qiang Cheng. TSCMamba: Mamba meets multi-view learning for time series classification. Information Fusion, 120:103079, 2025. Önder Akacik and Mark Hoogendoorn. ModernTCN revisited: A critical look at the experimental setup in general time series analysis. Transactions on Machine Learning Research, 2025. 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. URL https://arxiv.org/abs/1811.00075. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate, 2016. URL https://arxiv.org/abs/1409.0473. Accepted 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, 2018. URL https://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, 1994. 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. In Proceedings of the 41st International Conference on Machine Learning, volume 235, pp. 10041–10071, 2024. Angus Dempster, François Petitjean, and Geoffrey I Webb. ROCKET: Exceptionally fast and accurate time series classification using random convolutional kernels. Data Mining and Knowledge Discovery, 34(5):1454–1495, 2020. 11

Published as a conference paper at ICLR 2026

Angus Dempster, Daniel F. Schmidt, and Geoffrey I. Webb. Minirocket: A very fast (almost) deterministic transform for time series classification. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pp. 248–257, 2021. Angus Dempster, Daniel F Schmidt, and Geoffrey I Webb. Hydra: Competing convolutional kernels for fast and accurate time series classification. Data Mining and Knowledge Discovery, 37:1779– 1805, 2023. Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, and Xiaoli Li. TSLANet: Rethinking Transformers for time series representation learning. In Proceedings of the 41st International Conference on Machine Learning, volume 235, pp. 12409–12428, 2024. Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. In The First Conference on Language Modeling, 2024. Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. In Advances in Neural Information Processing Systems, pp. 572–585, 2021. Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2022. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016. Weizhe Hua, Zihang Dai, Hanxiao Liu, and Quoc Le. Transformer quality in linear time. In Proceedings of the 39th International Conference on Machine Learning, volume 162, pp. 9099–9117, 2022. Hassan Ismail Fawaz, Benjamin Lucas, Germain Forestier, Charlotte Pelletier, Daniel F Schmidt, Jonathan Weber, Geoffrey I Webb, Lhassane Idoumghar, Pierre-Alain Muller, and François Petitjean. InceptionTime: Finding alexnet for time series classification. Data Mining and Knowledge Discovery, 34(6):1936–1962, 2020. Jaeyoung Kim, Mostafa El-Khamy, and Jungwon Lee. Residual LSTM: Design of a deep recurrent architecture for distant speech recognition. In Interspeech 2017, pp. 1591–1595, 2017. Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In International Conference on Learning Representations, 2020. Prabhat Kumar and S. Suresh. FLAAP: An open human activity recognition (HAR) dataset for learning and finding the associated activity patterns. Procedia Computer Science, 212:64–73, 2022. Barak Lenz, Opher Lieber, Alan Arazi, Amir Bergman, Avshalom Manevich, Barak Peleg, Ben Aviram, Chen Almagor, Clara Fridman, Dan Padnos, Daniel Gissin, Daniel Jannai, Dor Muhlgay, Dor Zimberg, Edden M. Gerber, Elad Dolev, Eran Krakovsky, Erez Safahi, Erez Schwartz, Gal Cohen, Gal Shachaf, Haim Rozenblum, Hofit Bata, Ido Blass, Inbal Magar, Itay Dalmedigos, Jhonathan Osin, Julie Fadlon, Maria Rozman, Matan Danos, Michael Gokhman, Mor Zusman, Naama Gidron, Nir Ratner, Noam Gat, Noam Rozen, Oded Fried, Ohad Leshno, Omer Antverg, Omri Abend, Or Dagan, Orit Cohavi, Raz Alon, Ro’i Belson, Roi Cohen, Rom Gilad, Roman Glozman, Shahar Lev, Shai Shalev-Shwartz, Shaked Haim Meirom, Tal Delbari, Tal Ness, Tomer Asida, Tom Ben Gal, Tom Braude, Uriya Pumerantz, Josh Cohen, Yonatan Belinkov, Yuval Globerson, Yuval Peleg Levy, and Yoav Shoham. Jamba: Hybrid transformer-mamba language models. In The Thirteenth International Conference on Learning Representations, 2025. Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part XXVI, pp. 237–255, 2024. 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, 2023. URL https://arxiv.org/abs/2302. 04501. Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. In International Conference on Learning Representations, 2020. Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. iTransformer: Inverted Transformers are effective for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024. Donghao Luo and Xue Wang. ModernTCN: A modern pure convolution structure for general time series analysis. In The Twelfth International Conference on Learning Representations, 2024. Stéphane Mallat. A wavelet tour of signal processing. Elsevier, 1999. Leland McInnes, John Healy, and James Melville. UMAP: Uniform manifold approximation and projection for dimension reduction, 2020. URL https://arxiv.org/abs/1802.03426. Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. In The Eleventh International Conference on Learning Representations, 2023a. Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. In The Eleventh International Conference on Learning Representations, 2023b. Matthew Middlehurst, James Large, and Anthony Bagnall. The canonical interval forest (cif) classifier for time series classification. In 2020 IEEE international conference on big data (big data), pp. 188–195. IEEE, 2020. Matthew Middlehurst, James Large, and Anthony Bagnall. HIVE-COTE 2.0: a new meta ensemble for time series classification. Machine Learning, 110:3211–3243, 2021. Matthew Middlehurst, Ali Ismail-Fawaz, Antoine Guillaume, Christopher Holder, David GuijoRubio, Guzal Bulatova, Leonidas Tsaprounis, Lukasz Mentel, Martin Walter, Patrick Schäfer, et al. aeon: a python toolkit for learning from time series. Journal of Machine Learning Research, 25(289):1–10, 2024. Andreas Miltiadous, Katerina D Tzimourta, Theodora Afrantou, Panagiotis Ioannidis, Nikolaos Grigoriadis, Dimitrios G Tsalikakis, Pantelis Angelidis, Markos G Tsipouras, Euripidis Glavas, Nikolaos Giannakeas, et al. A dataset of scalp EEG recordings of alzheimer’s disease, frontotemporal dementia and healthy subjects from routine EEG. Data, 8(6):95, 2023. Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with Transformers. In The Eleventh International Conference on Learning Representations, 2023. 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. Francesco Spinnato, Riccardo Guidotti, Anna Monreale, and Mirco Nanni. Fast, interpretable and deterministic time series classification with a bag-of-receptive-fields. IEEE Access, 2024. state-spaces. mamba. https://github.com/state-spaces/mamba, 2025. Accessed: 8 August 2025. Sabera J Talukder, Yisong Yue, and Georgia Gkioxari. TOTEM: TOkenized time series EMbeddings for general time series analysis. Transactions on Machine Learning Research, 2024. ISSN 28358856. 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, 2022. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 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. In The Thirteenth International Conference on Learning Representations, 2025a. Yihe Wang, Nan Huang, Taida Li, Yujun Yan, and Xiang Zhang. Medformer: A multi-granularity patching transformer for medical time-series classification. In Advances in Neural Information Processing Systems, volume 37, pp. 36314–36341, 2024a. Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Mingsheng Long, and Jianmin Wang. Deep time series models: A comprehensive survey and benchmark, 2024b. URL https://arxiv. org/abs/2407.13278. Zihan Wang, Fanheng Kong, Shi Feng, Ming Wang, Xiaocui Yang, Han Zhao, Daling Wang, and Yifei Zhang. Is Mamba effective for time series forecasting? Neurocomputing, 619:129178, 2025b. Yunshi Wen, Tengfei Ma, Ronny Luss, Debarun Bhattacharjya, Achille Fokoue, and Anak Agung Julius. Shedding light on time series classification using interpretability gated networks. In The Thirteenth International Conference on Learning Representations, 2025. Frank Wilcoxon. Individual comparisons by ranking methods. Biometrics bulletin, 1(6):80–83, 1945. Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. ETSformer: Exponential smoothing Transformers for time-series forecasting, 2022. URL https://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. In The Eleventh International Conference on Learning Representations, 2023. Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, Yunhai Tong, and Bixiong Xu. TS2Vec: Towards universal representation of time series. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp. 8980–8987, 2022. Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are Transformers effective for time series forecasting? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 11121–11128, 2023. T Zhang, Y Zhang, W Cao, J Bian, X Yi, S Zheng, and J Li. Less Is More: Fast multivariate time series forecasting with light sampling-oriented MLP structures, 2022. URL https://arxiv. org/abs/2207.01186. Yunhao Zhang and Junchi Yan. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. In The Eleventh International Conference on Learning Representations, 2023. Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. FEDformer: Frequency enhanced decomposed Transformer for long-term series forecasting. In Proceedings of the 39th International Conference on Machine Learning, volume 162, pp. 27268–27286, 2022. Tian Zhou, Peisong Niu, Xue Wang, Liang Sun, and Rong Jin. One Fits All: Power general time series analysis by pretrained LM. In Advances in Neural Information Processing Systems, volume 36, pp. 43322–43355, 2023.

Published as a conference paper at ICLR 2026

E XPERIMENTAL SETUP

Environment All 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 the aeon toolkit (Middlehurst et al., 2024), and all deep models except TSLANet were integrated into the Time-Series Library (TSLib) (Wu et al., 2023; Wang et al., 2024b). For models supported in TSLib, we used the provided code; for others, we adapted the authors’ official implementations. All source code is available in our GitHub repository (see reproducibility statement). Dataset The UEA archive (Bagnall et al., 2018) provides 30 multivariate TSC datasets with diverse sample sizes, input dimensions, lengths, and class counts (Table 4). As TSLib has become a widely used framework, a subset of 10 datasets (EC, FD, HW, HB, JV, PEMS, SRS1, SRS2, SAD, and UW) is commonly used in recent TSC benchmarking practices. Table 4: Summary of the 30 UEA datasets used in our experiments Dataset

Code

Train Size

Test Size

Length

Variables

Classes

EthanolConcentration FaceDetection Handwriting Heartbeat JapaneseVowels PEMS-SF SelfRegulationSCP1 SelfRegulationSCP2 SpokenArabicDigits UWaveGestureLibrary ArticularyWordRecognition AtrialFibrillation BasicMotions CharacterTrajectories Cricket DuckDuckGeese EigenWorms Epilepsy ERing FingerMovements HandMovementDirection InsectWingbeat Libras LSST MotorImagery NATOPS PenDigits PhonemeSpectra RacketSports StandWalkJump

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ

261 5890 150 204 270 267 268 200 6599 120 275 15 40 1422 108 50 128 137 30 316 160 25000 180 2459 278 180 7494 3315 151 12

263 3524 850 205 370 173 293 180 2199 320 300 15 40 1436 72 50 131 138 270 100 74 25000 180 2466 100 180 3498 3353 152 15

1751 62 152 405 7–29 144 896 1152 4–93 315 144 640 100 60–182 1197 270 17984 206 65 50 400 2–22 45 36 3000 51 8 217 30 2500

3 144 3 61 12 963 6 7 13 3 9 2 6 3 6 1345 6 3 4 28 10 200 2 6 64 24 2 11 6 4

4 2 26 2 9 7 2 2 10 8 25 3 4 20 12 5 5 4 6 2 4 10 15 14 2 6 10 39 4 3

Metrics The evaluation metrics reported in our figures and tables are summarized as follows: • avg. acc: Average accuracy (%) across all UEA datasets. • avg. acc (10): Average accuracy (%) for 10 UEA datasets, following prior TSC practices. • avg. rank: Average rank of each model across all UEA datasets. • # of top-N : Count of datasets in which each model is in top-N . • # of win/draw: Count of datasets where proposed model achieved higher or the same accuracy. • # of loss/draw: Count of datasets where proposed model achieved lower or the same accuracy. • rank test: p-value of Wilcoxon signed-rank test (Wilcoxon, 1945). A value smaller than the significance level (0.05) indicates that proposed model ranked greater than the selected.

Published as a conference paper at ICLR 2026

H YPERPARAMETER SETTINGS

B.1

BASIC HYPERPARAMETER SETTINGS

We standardized experimental hyperparameter settings across all models to focus on model tuning, where most were chosen as common values across prior UEA benchmarks and TSLib: • Batch size: 16 (some experiments used smaller batch size as described in appendix C.1.) • Learning rate: 0.001 • Learning rate scheduler: none • Optimizer: RAdam (Liu et al., 2020) • 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

M ODEL HYPERPARAMETER SETTINGS FOR GRID SEARCH

We primarily considered the hyperparameter settings that were tested in either the original papers or the source code scripts, and we also referred to the hyperparameters that were set in TSLib scripts. For TSF-origin and foundation models that do not provide classification scripts, we tend to choose the same or smaller values than in forecasting scripts since we empirically found that many UEA datasets require compact model sizes compared to TSF datasets. Over 200 candidate settings were evaluated per model–dataset pair, except for some models that did not require significant extension. The hyperparameters explored are as below: • Non-DL – DTWD : fixed implementations (no learnable hyperparameters). – ROCKET, HC2, Hydra, MR+Hydra: 3 combinations * seed: 0, 1, 2 * Other parameters are fixed to the default optimal setting of the original papers since the models themselves have either large number (1e-4) of kernels or ensemble mechanism. • MLP-based models – DLinear: 15 combinations * moving avg: 0.5%, 1%, 2%, 3%, 4%, 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50% of sequence length * Since DLinear only has one model parameter to be tuned, we didn’t over-expand the number of hyperparameter combinations beyond 200. – LightTS: 100 combinations * d model: 32, 64, 128, 256, 512 * chunk size: 1/21 – 1/2 of sequence length * Since LightTS only has two model parameters to be tuned, we didn’t over-expand the number of hyperparameter combinations beyond 200. – 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) if d ff is 0, else 1 (True) * down sampling window: 0, 1%, 2%, 3%, 5%, 7.5%, 10%, 12.5% of sequence length * fac T: 0 (False) if down sampling window is 0, else 1 (True) * use norm: 1 (True) 16

Published as a conference paper at ICLR 2026

• CNN-based models – TimesNet: 252 combinations * e layers: 2, 3, 4 * (d model, d ff): (8,16), (8,32), (8,64), (16,16), (16,32), (16,64), (32,32), (32,64), (32,128), (64,64), (64,128), (64,256), (128,128), (128,256) * top k: 1, 2, 3 * num kernels: 4, 6 – ModernTCN: 252 combinations * ffn ratio: 1, 2, 4 * patch size: 2.5%, 5%, 7.5%, 10%, 15%, 20%, 25% of sequence length * patch stride: 50% of patch size * (num blocks, large size, small size, dims): (“1”, “13”, “5”, “32”), (“1”, “13”, “5”, “64”), (“1”, “13”, “5”, “128”), (“1”, “13”, “5”, “256”), (“1 1”, “9 9”, “5 5”, “32 64”), (“1 1”, “9 9”, “5 5”, “64 128”), (“1 1”, “9 9”, “5 5”, “128 256”), (“1 1”, “13 13”, “5 5”, “32 64”), (“1 1”, “13 13”, “5 5”, “64 128”), (“1 1”, “13 13”, “5 5”, “128 256”), (“1 1 1”, “9 9 9”, “5 5 5”, “32 64 128”), (“1 1 1”, “13 13 13”, “5 5 5”, “32 64 128”) – TSLANet: 252 combinations * depth: 1, 2, 3 * emb dim: 32, 64, 128, 256 * mlp ratio: 1, 2, 3 * patch size: 2.5%, 5%, 7.5%, 10%, 15%, 20%, 25% of sequence length * patch stride: 50% of patch size * masking ratio: 0.4 * ICB: 1 (True) * ASB: 1 (True) * adaptive filter: 1 (True) * load from pretrained: 1 (True) * pretrain lr: 0.001 * pretrain epoch: 50 – TimeMixer++: 216 combinations * down sampling method: “conv” * down sampling layers: 1, 2, 3 * down sampling window: 2 * num kernels: 6 * e layers: 1, 2, 3, 4 * d model: 16, 32, 64 * d ff: 16, 32, 64 * n heads: 8 * top k: 2, 3

Published as a conference paper at ICLR 2026

• Transformer-based models – FEDformer: 250 combinations * moving avg: 0.5%, 1%, 2%, 3%, 5%, 10%, 20%, 30%, 40%, 50% of sequence length * e layers: 2 * d model: 32, 64, 128, 256, 512 * d ff: 128, 256, 512, 1024, 2048 * n heads: 8 – ETSformer: 240 combinations * e layers: 1, 2 * d model: 32, 64, 128, 256, 512 * d ff: 64, 128, 256, 512, 1024, 2048 * n heads: 8 * top k: 1, 2, 3, 4 – Crossformer: 288 combinations * e layers: 1, 2, 3 * d model: 32, 64, 128, 256 * d ff: 64, 128, 256, 512 * n heads: 4 * factor: 3, 10 * seg len: 6, 12, 24 (change the code of TSLib to modify the fixed value) – PatchTST: 252 combinations * e layers: 1, 2, 3 * d model: 16, 32, 64, 128 * d ff: 64, 128, 256 * n heads: 4 if d model ∈ {16, 32} else 16 * patch size: 2.5%, 5%, 7.5%, 10%, 15%, 20%, 25% of sequence length * patch stride: 50% of patch size – GPT4TS: 252 combinations * e layers: 3, 4, 5, 6 * d model: 768 (fixed since the model use pretrained GPT-2) * d ff: 768 (fixed since the model use pretrained GPT-2) * patch size: 1/25 – 1/5 of sequence length * patch stride: 25%, 50%, 100% of patch size – iTransformer: 240 combinations * e layers: 1, 2, 3, 4 * d model: 64, 128, 256, 512, 1024, 2048 * d ff: 64, 128, 256, 512, 1024 * n heads: 8 * factor: 1, 3

Published as a conference paper at ICLR 2026

• Shape-based model – InterpGN: 243 combinations * dnn type: “FCN” * num shaplet: 5, 10, 15 * lambda div: 0, 0.1, 1.0 * lambda reg: 0, 0.1, 1.0 * epsilon: 0.5, 1.0, 2.0 * gating value: 0.5, 0.75, 1.0 * distance func: “euclidean” * memory efficient: False * sbm cls: “linear” • Mamba-based models – TSCMamba: 288 combinations * d model: 64, 128, 256 * e layers: 1, 2 * expand: 1, 2 * d conv: 2, 4 * d ff (d state): 32, 64, 128 * no rocket: 0 (False) * half rocket: 0 (False), 1 (True) * additive fusion: 0 (multiplicative fusion), 1 (additive fusion) * max pooling: 0 (avg pooling), 1 (max pooling) * channel token mixing: 0 (False) * only forward scan: 0 (False) * flip dir: 2 (vertical flip) * reverse flip: 0 (False) * patch size: 8 * patch stride: 8 * rescale size: 64 * variation: 64 * initial focus: 1.0 – MambaSL (ours): 240 combinations * d model (dm ): 32, 64, 128, 256, 512, 1024 * d ff (ds ): 1, 2, 4, 8, 16 * tv dt (θ∆ ): 0 (False), 1 (True) * tv B (θB ): 0 (False), 1 (True) * tv C (θC ): 0 (False), 1 (True) * use D: 0 (False) * expand: 1 * d conv: 4

Published as a conference paper at ICLR 2026

A DDITIONAL E XPERIMENTAL R ESULTS

C.1

F ULL RESULTS OF TSC MODELS ON THE UEA BENCHMARK

Table 5 presents the full experimental results corresponding to Figure 4. As mentioned in appendix A, some experiments could not be run on the GTX 1080 Ti GPU with batch size 16. Footnotes in the table indicate such experiments. We only used the A100 GPU for cases that could not be run on the 1080 Ti with batch size 4. Also, note that only one hyperparameter combination was tested in our limited resource environment for TSCMamba (Ahamed & Cheng, 2025)–IW pair, 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 in bold and underline, respectively. (a) Classification results from DTWD to ModernTCN. non-DL HC2 (2021)

MLP-based Hydra (2023)

MR+ Hydra (2023)

ROCKET (2020)

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ

32.319 52.866 60.706 71.707 95.946 71.098 77.474 53.889 97.226 90.313 98.667 20.000 97.500 98.816 100.000 58.000 61.832 96.377 91.481 53.000 18.919 48.692 87.222 55.109 50.000 88.333 97.713 15.121 80.263 20.000

44.487 52.852 54.373 60.076 61.606 61.379 60.982 61.294 59.294 56.000 56.353 53.529 75.610 78.537 76.585 77.073 83.784 98.378 97.838 98.378 84.393 97.688 98.266 82.659 85.666 89.078 86.689 94.881 55.556 54.444 58.889 55.000 99.227 99.136 99.136 99.045 94.063 94.063 94.375 94.063 99.333 99.333 99.333 99.333 6.667 20.000 26.667 6.667 100.000 100.000 100.000 100.000 99.304 99.304 99.164 99.373 100.000 100.000 100.000 98.611 52.000 70.000 68.000 60.000 91.603 97.710 98.473 98.473 99.275 100.000 100.000 100.000 98.889 98.889 98.889 99.259 57.000 61.000 55.000 60.000 51.351 48.649 51.351 36.486 38.524 66.796 67.052 65.744 90.556 94.444 93.333 93.333 64.031 66.383 66.504 64.274 54.000 50.000 54.000 51.000 89.444 91.111 90.000 93.333 98.228 97.656 97.856 97.284 27.826 33.015 33.164 33.642 90.789 90.132 91.447 90.132 46.667 40.000 40.000 66.667

31.179 32.319 30.798 33.460 32.319 69.495 68.473 68.530 70.148 66.998 23.765 25.059 19.647 37.529 30.118 76.585 77.561 79.024 83.902 78.049 96.486 96.757 95.946 98.649 98.649 82.659 87.861 82.081 87.861 86.705A 92.833 92.833 78.840 92.150 93.515 57.222 59.444 58.333 60.000 60.000 96.953 98.226 99.454 99.545 99.500 82.813 89.063 83.125 88.438 88.438 98.667 98.667 97.667 98.333 99.000 53.333 73.333 60.000 66.667 73.333 87.500 100.000 100.000 100.000 100.000 98.120 98.955 97.981 98.747 99.304 91.667 93.056 97.222 97.222 97.222 8 66.000 56.000 68.000 68.000 66.000A 42.748A 54.1988 66.412 63.359 64.122 65.217 97.101 66.667 96.377 96.377 91.852 94.444 77.778 95.556 97.407 62.000 64.000 63.000 66.000 67.000 66.216 66.216 70.270 72.973 66.216 18.996 71.676 72.424 64.024 75.588 70.556 89.444 34.444 88.333 90.000 34.023 41.525 34.185 44.120 43.593 64.000 68.000 66.000 68.000 67.000 96.111 97.222 95.000 98.333 95.556 92.481 96.998 46.941 98.914 98.513 6.651 16.224 5.428 14.524 13.749 78.947 81.579 86.842 89.474 84.211 53.333 73.333 66.667 66.667 80.000

avg. acc (10) avg. acc avg. rank # of top-1 # of top-5 # of top-10

70.354 68.020 16.03 1 2 5

74.368 73.306 11.20 3 10 15

78.156 76.866 8.90 4 13 19

78.349 77.124 8.83 7 13 17

77.600 76.320 9.40 8 12 16

70.999 68.247 16.80 0 0 5

72.760 75.308 11.83 1 6 11

69.578 69.001 13.90 1 3 9

75.168 76.921 8.67 3 9 17

73.429 76.938 9.13 2 8 16

# of win/draw # of lose/draw rank test

29 2 0.000

24 9 0.001

20 14 0.056

20 13 0.034

21 11 0.015

28 3 0.000

25 7 0.000

28 4 0.000

24 8 0.004

24 8 0.001

LightTS (2022)

TimesNet (2023)

Modern TCN (2024)

DTWD (1994)

All experiments were run on 1080 Ti with batch size 16 except below: 8 The experiments were run with batch size 8 4 The experiments were run with batch size 4 A The experiments were run on A100 GPU

DLinear (2023)

CNN-based MTSMixer (2023)

Published as a conference paper at ICLR 2026

Table 5: (Continued.) Classification accuracy (%) of TSC models on 30 datasets from the UEA archives. The best and the second-best are highlighted in bold and underline, respectively. (b) Classification results from TSLANet to MambaSL. “*” indicates “former” in Transformer-based models. CNN-based

shapebased

Transformer-based

Time TSLANet FED* Mixer++ (2024) (2022) (2025a)

ETS* (2022)

Mamba-based

TSC Cross* PatchTST GPT4TS iTrans* InterpGN MambaSL Mamba (2023) (2023) (2023) (2024) (2025) (proposed) (2025)

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ

31.939 34.601 33.840 33.080 43.346 66.969 69.665 69.892 68.331 69.637 62.000 33.647 33.412 35.647 30.353 79.512 80.488 78.049 78.537 79.024 98.919 99.189 97.568 98.378 98.378 89.017 89.017 89.017 87.861 89.5958 89.761 92.150 75.085 92.150 93.857 63.889 60.556 59.444 60.000 61.667 98.045 99.545 99.454 99.000 98.863 94.375 90.625 79.688 89.688 88.438 99.333 99.000 92.333 98.667 98.667 66.667 66.667 73.333 80.000 73.333 100.000 100.000 95.000 100.000 100.000 99.164 99.373 97.284 98.468 98.955 91.667 98.611 90.278 98.611 97.222 74.000 74.000 42.000 42.000 72.0004 81.679 61.069A 63.3594 70.9924 58.0154 100.000 94.928 92.754 92.754 95.652 85.926 96.296 95.926 96.296 97.407 68.000 67.000 72.000 66.000 64.000 48.649 72.973 71.622 63.514 66.216 10.012 64.192 65.244 64.868 73.712 95.556 88.333 93.333 90.000 90.556 51.014 44.444 40.998 41.241 41.768 71.000 67.000 65.000 67.0008 68.0004 99.444 96.111 96.111 93.333 93.889 99.028 98.599 98.599 98.542 98.513 4.921 15.508 11.691 14.316 14.256 94.079 91.447 84.868 85.526 87.500 73.333 73.333 66.667 66.6678 80.000

32.700 32.319 68.076 68.615 29.529 34.118 73.659 79.024 97.027 98.108 89.0174 88.439 89.078 93.174 61.667 60.556 98.818 99.591 89.375 89.063 98.667 98.667 80.000 66.667 85.000 100.000 98.398 99.373 97.222 97.222 40.0004 66.000 57.252 66.412 98.551 89.855 96.667 96.296 63.000 66.000 63.514 70.270 60.936 61.396 81.667 87.778 56.732 38.078 69.000 67.000 80.556 96.111 98.485 98.513 15.091 13.123 85.526 86.842 86.667 73.333

avg. acc (10) avg. acc avg. rank # of top-1 # of top-5 # of top-10

77.443 74.948 71.545 74.267 75.316 76.263 77.279 74.128 75.716 77.427 7.80 7.03 12.33 10.63 8.50 7 2 1 2 1 16 12 6 2 12 21 24 10 13 17

72.895 74.301 74.120 74.729 76.065 74.797 12.60 9.50 11.83 2 1 0 4 6 4 8 18 10

75.482 75.933 77.700 78.405 7.87 7.10 7 5 14 12 20 25

# of win/draw # of lose/draw rank test

18 15 0.029

24 7 0.000

21 14 0.026

23 10 0.005

23 7 0.000

27 4 0.000

21 10 0.005

26 7 0.000

31.939 28.897 30.798 42.586 68.417 63.791 70.204 69.296 31.294 58.706 45.059 60.824 78.049 80.488 78.537 80.488 98.108 99.730 98.378 98.649 91.3298 86.127 90.173 85.549 93.857 90.444 92.833 92.491 58.889 58.333 62.222 65.000 99.318 99.864 95.816 99.955 90.000 88.438 95.313 93.438 99.000 99.000 99.333 99.333 66.667 46.667 66.667 53.333 97.500 100.000 100.000 100.000 98.886 99.791 99.513 99.721 97.222 100.000 100.000 100.000 60.0008 76.000A 50.000 70.000 59.542 82.443A 88.550 83.9694 80.435 97.826 99.275 97.826 94.815 87.778 97.037 93.704 64.000 67.000 66.000 71.000 58.108 45.946 60.811 70.270 73.436 67.180 68.244* 66.304 91.667 97.222 91.667 91.667 40.673 59.570 60.057 45.580 67.000 67.000A 67.000 69.000 88.889 98.889 92.222 98.889 98.428 99.342 98.513 99.257 11.542 33.075 23.800 30.331 81.579 91.447 90.789 92.763 73.333 60.000 73.333 73.333

25 7 0.000

All experiments were run on 1080 Ti with batch size 16 except below: 8 The experiments were run with batch size 8 4 The experiments were run with batch size 4 A The experiments were run on A100 GPU * Little hyperparameter search was performed since it took over two weeks conducting a single experiment.

20 15 0.039

78.827 79.819 5.30 5 18 27

Published as a conference paper at ICLR 2026

C.2

F ULL RESULTS OF MODEL ABLATION ON THE UEA BENCHMARK

Table 6 presents the complete ablation results corresponding to Table 1. For a clear comparison, we also performed the hyperparameter grid search for each ablated model. Model rankings were computed 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 hyperparameters (expand and d conv) and compared the results. As expected, performance improved with broader exploration, yet the model still underperformed on datasets such as EC and NATO, where the LTI setting remains advantageous. Table 6: Classification accuracy (%) of ablated models on all 30 datasets from the UEA archives. d model, d state, expand, and d conv are four basic hyperparameters that affect the model capacity of Mamba. The best and the second-best are highlighted in bold and underline, respectively. ablation w/o H1 w/o H2 w/o H3 kernel size ✓: k = 0.02L ✗: k = 3 modular SSM ✓: modular TV/TI ✗: TV only skip connection ✓: not use D ✗: use D aggregation ✓: adaptive pool. ✗: the others

additional proposed Mamba

w/o H4

w/ H2

w/o H2

✗

✓

✗

✓

✗

✓

✗

✓

✗

✓

✗

✓

✗

✓

✗ (full)

✗ (avg)

✗ (max)

✗ (last)

✓

✗ (full)

✗ (avg)

✓

d model d state expand d conv # of exp

32..1024 32..1024 32..1024 32..1024 32..1024 32..1024 32..1024

32..1024

32..1024 32..1024 32..1024

1..16 1 4 240

1..16 1 4 30

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ

33.080 32.319 39.544 34.601 32.700 41.825 33.080 69.296 68.587 69.211 70.006 70.034 69.012 66.827 60.235 54.000 58.000 47.294 59.882 45.529 56.706 80.976 79.024 80.000 80.000 80.488 80.976 78.537 98.649 98.108 98.649 98.649 98.108 98.919 98.919 85.549 82.659 87.283 90.751 82.659 85.549 89.595 92.150 89.761 91.809 93.174 91.468 90.102 87.713 61.111 61.111 60.556 61.111 60.000 63.333 61.111 99.955 99.773 99.909 99.727 99.909 99.864 99.818 93.125 91.250 92.188 89.063 93.750 89.375 86.250 99.333 99.333 99.000 99.000 99.333 99.000 97.000 73.333 46.667 53.333 66.667 53.333 66.667 66.667 100.000 100.000 100.000 100.000 100.000 100.000 100.000 99.582 99.443 99.513 99.095 99.791 99.234 99.582 100.000 100.000 100.000 98.611 100.000 100.000 100.000 70.000 66.000 74.000 64.000 72.000 76.000 80.000 85.496 83.206 84.733 67.176 83.969 84.733 51.908 97.826 97.826 97.101 96.377 97.826 98.551 97.826 93.704 91.852 94.444 95.926 95.185 91.852 92.222 71.000 64.000 65.000 69.000 64.000 67.000 68.000 72.973 64.865 68.919 68.919 68.919 56.757 50.000 66.304 65.952 66.220 65.124 66.084 66.392 66.584 91.667 91.667 90.556 86.111 91.111 88.889 88.889 45.580 43.593 46.513 43.390 45.174 43.593 42.701 65.000 65.000 70.000 70.000 65.000 66.000 65.000 98.889 97.222 98.889 97.778 98.889 98.889 98.889 99.257 99.142 99.371 99.114 99.428 99.085 98.999 29.615 29.764 30.063 16.880 31.792 27.528 27.140 92.763 90.132 93.421 90.789 90.789 94.737 88.816 80.000 60.000 66.667 73.333 66.667 66.667 66.667

42.586 69.296 60.824 80.488 98.649 85.549 92.491 65.000 99.955 93.438 99.333 53.333 100.000 99.721 100.000 70.000 83.969 97.826 93.704 71.000 70.270 66.304 91.667 45.580 69.000 98.889 99.257 30.331 92.763 73.333

30.418 34.221 33.460 68.785 69.154 69.495 30.941 60.471 56.000 77.073 80.976 80.976 97.297 97.838 98.919 93.064 83.815 85.549 92.833 91.126 91.809 58.889 60.556 62.222 99.181 99.955 99.818 84.375 92.813 93.438 97.667 99.000 99.333 66.667 46.667 60.000 97.500 100.000 100.000 98.816 99.791 99.582 97.222 100.000 100.000 52.000 72.000 74.000 67.176 75.573 85.496 92.754 97.101 97.826 94.444 96.296 94.815 64.000 64.000 68.000 52.703 74.324 68.919 62.684 66.396 66.756 82.778 92.778 92.222 37.307 43.917 44.282 65.000 63.000 66.000 95.556 98.333 98.333 98.799 99.314 99.314 13.809 31.226 29.764 84.211 94.079 90.789 73.333 53.333 73.333

avg. acc (10) avg. acc avg. rank

77.413 80.215 2.53

75.659 77.075 5.93

77.715 79.163 3.53

76.438 77.722 4.90

76.900 78.610 3.60

76.448 78.535 3.90

75.856 76.848 4.87

78.827 73.286 79.819 74.243 2.40

77.092 77.935

77.169 79.348

# of win/draw # of lose/draw rank test

25 22 0.217

30 5 0.000

22 14 0.071

24 10 0.003

22 15 0.044

22 12 0.004

25 9 0.001

26 5 0.000

20 13 0.011

20 17 0.106

1..16 1 4 30

1..16 1 4 240

1..16 1,2,4 1,2,4 270

Published as a conference paper at ICLR 2026

C.3

C OMPARISON WITH PREVIOUSLY REPORTED RESULTS

Although we compared various hyperparameter settings against the original paper to ensure as fair a comparison as possible, our results may not represent those in the original paper due to some differences in basic hyperparameter settings (e.g. fixing the learning rate to 0.001). Therefore, we compared our results with previously reported results, and the results are presented in Table 7. Among the baseline models, only papers that directly evaluated the classification results are used and mentioned in the table. If multiple papers reported results for a particular model–dataset pair, the highest value was used for comparison. For Hydra and MR+Hydra, the results reported in BORF (Spinnato et al., 2024) 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 higher than “report ≥ ours” indicates that existing reports likely underestimated the model. • report ≥ ours: Count of datasets on which our experimental results of grid search achieved lower or equal accuracy compared to the reported results. A value higher than “report ≤ ours” indicates that our experiments likely underestimated the model. • acc difference: Average difference between our experimental accuracy and the reported accuracy, calculated only for datasets with paired results. A positive number indicates that our experiments achieved higher accuracy, while a negative number indicates the opposite. Table 7: Comparison of our experimental results and previous reported results. (a) Comparison results on TSC-specific models. Distance DTWD (1994) TimesNet ours

TSC-specific ROCKET (2020) TimesNet TSLANet TSCMamba

ours

HC2 (2021) original

ours

Hydra (2023) BORF

ours

MR+Hydra (2023) BORF

ours

InterpGN (2025) original

ours

TSCMamba (2025) original

ours

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ

32.281 32.319 45.209 44.487 79.087 52.852 54.791 54.373 58.897 60.076 30.418 28.897 62.015 30.798 52.900 52.866 64.699 61.606 71.348 61.379 52.301 60.982 61.901 61.294 66.402 63.791 69.401 70.204 28.600 60.706 58.800 59.294 56.341 56.000 46.600 56.353 51.400 53.529 61.882 58.706 53.306 45.059 71.707 71.707 75.610 75.610 72.878 78.537 76.098 76.585 73.220 77.073 79.512 80.488 76.585 78.537 94.892 95.946 96.189 83.784 98.378 97.811 97.838 97.811 98.378 99.189 99.730 97.000 98.378 71.098 71.098 75.087 84.393 99.827 97.688 80.289 98.266 77.514 82.659 88.439 86.127 90.173 90.173 77.713 77.474 90.785 85.666 87.884 89.078 86.689 86.689 95.188 94.881 92.491 90.444 92.491 92.833 53.889 53.889 54.444 55.556 50.444 54.444 58.889 58.889 56.722 55.000 57.778 58.333 66.722 62.222 96.298 97.226 99.200 99.227 99.136 98.899 99.136 99.300 99.045 99.818 99.864 99.000 95.816 90.313 90.313 94.406 94.063 94.875 94.063 90.906 94.375 94.094 94.063 91.875 88.438 93.813 95.313 98.667 99.333 99.333 99.567 99.333 99.000 99.333 99.333 99.333 99.667 99.000 97.000 99.333 20.000 20.000 6.667 28.000 20.000 13.333 26.667 6.667 6.667 33.333 46.667 67.333 66.667 97.500 100.000 100.000 99.000 100.000 100.000 100.000 100.000 100.000 100.000 100.000 100.000 100.000 98.816 98.753 99.304 99.304 98.600 99.164 99.199 99.373 99.791 99.791 98.997 99.513 100.000 98.611 100.000 100.000 100.000 98.611 100.000 98.611 98.611 100.000 100.000 98.611 100.000 58.000 52.000 52.000 49.800 70.000 68.000 68.000 64.000 60.000 68.000 76.000 62.000 50.000 61.832 78.626 91.603 89.466 97.710 98.473 98.473 96.183 98.473 83.969 82.443 87.023 88.550 96.377 98.551 99.275 99.783 100.000 100.000 100.000 100.000 100.000 99.275 97.826 97.101 99.275 91.481 94.074 98.889 98.519 98.889 98.519 98.889 98.889 99.259 95.185 87.778 91.519 97.037 53.000 61.000 57.000 55.200 61.000 55.000 55.000 56.000 60.000 64.000 67.000 69.000 66.000 18.919 50.000 51.351 39.730 48.649 21.622 51.351 36.486 36.486 48.649 45.946 71.622 60.811 48.692 10.000 38.524 66.796 67.900 67.052 66.700 65.744 66.752 67.180 14.300 68.244 87.222 83.889 90.556 92.667 94.444 93.278 93.333 92.778 93.333 97.222 97.222 90.000 91.667 55.109 54.100 64.031 63.694 66.383 60.702 66.504 63.601 64.274 60.665 59.570 60.300 60.057 50.000 53.000 54.000 53.200 50.000 50.000 54.000 48.000 51.000 64.000 67.000 62.000 67.000 88.333 83.333 89.444 89.222 91.111 87.222 90.000 91.722 93.333 98.889 98.889 94.444 92.222 97.713 97.341 98.228 99.557 97.656 96.801 97.856 96.801 97.284 99.142 99.342 98.542 98.513 15.121 17.599 27.826 29.427 33.015 32.401 33.164 35.401 33.642 32.180 33.075 24.659 23.800 80.263 86.184 90.789 93.026 90.132 91.382 91.447 90.789 90.132 90.789 91.447 91.447 90.789 20.000 46.667 46.667 44.000 40.000 40.000 40.000 60.000 66.667 53.333 60.000 73.333 73.333

report ≤ ours report ≥ ours acc difference

8 6 3.385

23 12 2.056

15 12 0.224

28 10 3.320

21 15 0.747

18 17 0.278

17 16 0.080

Published as a conference paper at ICLR 2026

Table 7: (Continued.) Comparison of our experimental results and previous reported results. (b) Comparison results on foundation models. Foundation TimesNet (2023) TSLANet original TSCMamba

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ

35.703 68.601 32.106 78.000 98.405 89.595 91.809 57.222 99.000 85.313

report ≤ ours report ≥ ours acc difference

8 2 1.593

ModernTCN (2024) ours

original

27.719 33.460 67.469 70.148 26.176 37.529 74.488 83.902 97.838 98.649 88.150 87.861 77.440 92.150 52.833 60.000 98.358 99.545 83.125 88.438 96.167 98.667 33.333 66.667 100.000 100.000 97.981 98.747 87.500 97.222 56.000 68.000 58.779 63.359 78.116 96.377 94.074 95.556 59.400 66.000 50.000 72.973 10.000 64.024 77.833 88.333 59.209 44.120 51.000 68.000 81.833 98.333 98.190 98.914 18.240 14.524 82.632 89.474 53.333 66.667

revisit

36.312 31.901 70.800 68.700 30.600 28.400 77.220 77.122 98.811 98.108 89.075 83.179 93.413 92.799 60.278 61.722 98.699 98.099 86.688 85.906

27 4 9.014

4 6 (0.761)

TSLANet (2024)

ours

original

32.319 30.418 66.998 66.771 30.118 57.882 78.049 77.561 98.649 99.189 86.705 83.815 93.515 91.809 60.000 61.667 99.500 99.909 88.438 91.250 98.667 99.000 73.333 40.000 100.000 100.000 99.304 97.222 98.611 66.000 64.122 96.377 98.551 97.407 67.000 61.000 66.216 52.703 75.588 10.000 90.000 92.778 43.593 66.338 67.000 62.000 95.556 95.556 98.513 98.939 13.749 17.751 84.211 90.789 80.000 46.667

8 2 0.835

19 8 2.161

TSCMamba

ours

TimeMixer++ (2025a)

GPT4TS (2023)

ours

TSLANet original TSCMamba

original

31.939 66.969 62.000 79.512 98.919 89.017 89.761 63.889 98.045 94.375 99.333 66.667 100.000 98.747 99.164 91.667 24.000 74.000 41.221 81.679 100.000 91.481 85.926 68.000 48.649 10.012 95.556 51.014 71.000 99.444 99.028 4.921 94.079 73.333

39.886 71.799 26.506 79.122 97.892 90.983 93.106 65.611 99.800 88.188

34.601 34.183 69.665 69.200 33.647 32.706 80.488 77.220 99.189 98.595 89.017 87.919 92.150 93.208 60.556 59.389 99.545 99.200 90.625 88.094 99.000 66.667 100.000 99.373 98.611 74.000 61.069 94.928 96.296 67.000 72.973 64.192 88.333 44.444 67.000 96.111 98.599 15.508 91.447 73.333

3 1 21.330

4 6 (0.341)

6 4 0.330

ours

25.475 32.319 65.579 68.615 3.765 34.118 36.585 79.024 98.108 98.108 87.283 88.439 91.468 93.174 51.667 60.556 99.363 99.591 84.375 89.063 93.333 98.667 33.333 66.667 92.500 100.000 98.259 99.373 8.333 97.222 50.000 66.000 48.092 66.412 85.507 89.855 95.926 96.296 57.000 66.000 18.919 70.270 10.000 61.396 79.444 87.778 46.391 38.078 50.000 67.000 91.667 96.111 97.742 98.513 3.012 13.123 76.974 86.842 33.333 73.333 29 2 15.617

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ report ≤ ours report ≥ ours acc difference

ours

LightTS (2022) TimesNet

ours

33.460 31.179 29.696 32.319 67.999 69.495 67.500 68.473 27.000 23.765 26.106 25.059 75.122 76.585 75.122 77.561 97.838 96.486 96.189 96.757 82.081 82.659 88.382 87.861 88.396 92.833 89.795 92.833 51.667 57.222 51.111 59.444 96.680 96.953 100.000 98.226 82.094 82.813 80.313 89.063 97.333 97.667 98.333 46.667 53.333 73.333 85.000 87.500 100.000 97.347 98.120 98.955 91.667 91.667 93.056 62.000 66.000 56.000 41.221 42.748 54.198 60.145 65.217 97.101 91.481 91.852 94.444 64.000 62.000 64.000 58.108 66.216 66.216 10.000 18.996 71.676 73.333 70.556 89.444 35.770 34.023 41.525 61.000 64.000 68.000 93.889 96.111 97.222 92.939 92.481 96.998 7.101 6.651 16.224 78.947 78.947 81.579 60.000 53.333 73.333 21 11 1.237

7 3 2.338

MTS-Mixer (2023) ModernTCN

ours

FEDformer (2022) ours

TimesNet

ETSformer (2022) TimesNet

ours

Crossformer (2023) ModernTCN TSLANet TSCMamba

ours

33.802 30.798 31.217 33.840 28.099 33.080 70.199 68.530 65.999 69.892 66.300 68.331 26.000 19.647 28.000 33.412 32.506 35.647 77.122 79.024 73.707 78.049 71.220 78.537 94.297 95.946 98.405 97.568 95.892 98.378 80.925 82.081 80.925 89.017 86.012 87.861 91.706 78.840 88.703 75.085 89.590 92.150 55.000 58.333 54.389 59.444 55.000 60.000 97.399 99.454 100.000 99.454 100.000 99.000 82.313 83.125 85.313 79.688 85.000 89.688 99.000 92.333 98.667 60.000 73.333 80.000 100.000 95.000 100.000 97.981 97.284 98.468 97.222 90.278 98.611 68.000 42.000 42.000 66.412 63.359 70.992 66.667 92.754 92.754 77.778 95.926 96.296 63.000 72.000 66.000 70.270 71.622 63.514 72.424 65.244 64.868 34.444 93.333 90.000 34.185 40.998 41.241 66.000 65.000 67.000 95.000 96.111 93.333 46.941 98.599 98.542 5.428 11.691 14.316 86.842 84.868 85.526 66.667 66.667 66.667

37.985 43.346 68.700 69.637 28.800 30.353 77.610 79.024 99.108 98.378 85.896 89.595 92.491 93.857 58.278 61.667 97.899 98.863 85.313 88.438 98.000 98.667 46.667 73.333 90.000 100.000 98.329 98.955 84.722 97.222 44.000 72.000 54.962 58.015 73.188 95.652 94.074 97.407 64.000 64.000 58.108 66.216 10.000 73.712 76.111 90.556 42.818 41.768 61.000 68.000 88.333 93.889 93.651 98.513 7.551 14.256 81.579 87.500 53.333 80.000

6 4 (1.298)

28 3 9.010

6 4 0.879

9 1 3.305

PatchTST (2023) ModernTCN TSLANet TSCMamba

ours TimeMixer++ ours

32.814 32.700 68.961 68.076 29.600 29.529 76.585 73.659 98.649 97.027 89.306 89.017 90.717 89.078 57.778 61.667 99.682 98.818 85.813 89.375 97.667 98.667 53.333 80.000 92.500 85.000 97.347 98.398 84.722 97.222 24.000 40.000 54.198 57.252 65.942 98.551 92.593 96.667 62.000 63.000 58.108 63.514 10.000 60.936 81.111 81.667 67.798 56.732 61.000 69.000 96.667 80.556 99.231 98.485 11.691 15.091 84.211 85.526 60.000 86.667 18 12 5.262

iTransformer (2024)

28.099 31.939 66.300 68.417 24.200 31.294 75.610 78.049 96.595 98.108 87.919 91.329 90.205 93.857 54.389 58.889 95.998 99.318 85.906 90.000 99.000 66.667 97.500 98.886 97.222 60.000 59.542 80.435 94.815 64.000 58.108 72.456 91.667 40.673 67.000 88.889 98.428 11.542 81.579 73.333 10 0 3.598

Published as a conference paper at ICLR 2026

C.4

V ISUALIZATION OF TIME VARIANCE ABLATION

average accuracy

Figure 9 shows visualizations of MambaSL’s classification accuracy results on the UEA benchmark, tested for various hyperparameter combinations along with TI/TV configurations. Among 240 combinations, 10 sets of hyperparameter combinations that showed high performance regardless of the TI/TV configurations were selected (8×10=80 in total) and depicted in box and line plots. Each line represents the same hyperparameter settings except for time variance configurations. 0.4 0.35 0.3

0.69 0.68 0.67 0.66

TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓

TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✗ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✗

configuration

0.6

(b) FaceDetection (FD) average accuracy

average accuracy

(a) EthanolConcentration (EC)

0.55

0.5

0.8

0.78

0.76

TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓

configuration

(d) Heartbeat (HB) average accuracy

average accuracy

0.98

0.97

0.85

0.8

TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✓

TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✓

configuration

0.92

(f) PEMS-SF (PEMS) average accuracy

average accuracy

(e) JapaneseVowels (JV)

0.9 0.88 0.86

0.65

0.6

0.55

TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓

TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗

configuration

(g) SelfRegulationSCP1 (SRS1)

(h) SelfRegulationSCP2 (SRS2) average accuracy

average accuracy

0.995

0.99

0.92

0.9

0.88

TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✓ TV-C:✗

TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓

configuration

(i) SpokenArabicDigits (SAD)

(j) UWaveGestureLibrary (UW)

Figure 9: Visualization of classification accuracy of MambaSL along TI/TV configurations on the UEA benchmark, in order of maximum accuracy.

average accuracy

Published as a conference paper at ICLR 2026

0.99 0.985 0.98

0.5 0.4 0.3 0.2

TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✓

TV-Δ:✗ TV-B:✗ TV-C:✗

configuration

average accuracy

1 0.5

TV-Δ:✗ TV-B:✗ TV-C:✓

TV-Δ:✗ TV-B:✓ TV-C:✗

TV-Δ:✗ TV-B:✓ TV-C:✓

TV-Δ:✓ TV-B:✗ TV-C:✗

TV-Δ:✓ TV-B:✗ TV-C:✓

TV-Δ:✓ TV-B:✓ TV-C:✗

TV-Δ:✓ TV-B:✓ TV-C:✓

configuration

average accuracy

TV-Δ:✓ TV-B:✓ TV-C:✓

(n) CharacterTrajectories (CT)

0.99

0.98

0.7 0.65 0.6 0.55 0.5

TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✗

TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✓ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓

configuration

(p) DuckDuckGeese (DDG)

0.85

0.98

average accuracy

TV-Δ:✓ TV-B:✗ TV-C:✓

TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓

(o) Cricket (CR)

0.8 0.75 0.7

0.96 0.94 0.92 0.9

TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✓ configuration

(r) Epilepsy (EP) average accuracy

0.92 0.9 0.88

TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✓ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✓ configuration

(q) EigenWorms (EW) average accuracy

TV-Δ:✗ TV-B:✗ TV-C:✓

0.99

configuration

0.86

TV-Δ:✗ TV-B:✓ TV-C:✗

0.995

(m) BasicMotions (BM)

0.97

TV-Δ:✓ TV-B:✗ TV-C:✗

(l) AtrialFibrillation (AF)

1.5

TV-Δ:✗ TV-B:✗ TV-C:✗

TV-Δ:✗ TV-B:✓ TV-C:✓

configuration

(k) ArticularyWordRecognition (AWR)

TV-Δ:✓ TV-B:✓ TV-C:✗

TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗

0.7 0.65 0.6 0.55

TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✓

configuration

(s) ERing (ER)

(t) FingerMovements (FM)

Figure 9: (Continued.) Visualization of classification accuracy of MambaSL along TI/TV configurations on the 30 UEA datasets, in order of maximum accuracy.

0.7

average accuracy

Published as a conference paper at ICLR 2026

0.6

0.5

TV-Δ:✓ TV-B:✗ TV-C:✓

TV-Δ:✗ TV-B:✗ TV-C:✗

TV-Δ:✓ TV-B:✗ TV-C:✗

TV-Δ:✗ TV-B:✓ TV-C:✓

TV-Δ:✓ TV-B:✓ TV-C:✗

TV-Δ:✗ TV-B:✗ TV-C:✓

TV-Δ:✓ TV-B:✓ TV-C:✓

0.66 0.65 0.64 0.63

TV-Δ:✗ TV-B:✓ TV-C:✗

TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✓ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✓

configuration

(u) HandMovementDirection (HMD)

(v) InsectWingbeat (IW) 0.46

average accuracy

0.92 0.9 0.88 0.86 0.84

0.44 0.42 0.4 0.38

TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✓

TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✗ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗

configuration

(x) LSST (LSST) average accuracy

average accuracy

(w) Libras (LIB)

0.6

0.55

0.5

0.98 0.96 0.94 0.92

TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✓

TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-B:✓ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓

configuration

(z) NATOPS (NATO) average accuracy

average accuracy

(y) MotorImagery (MI)

0.99

0.985

0.3 0.29 0.28 0.27 0.26

TV-Δ:✓ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✗ TV-B:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✗

TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✓ TV-C:✗

configuration

(ab) PhonemeSpectra (PS) average accuracy

average accuracy

(aa) PenDigits (PD)

0.9

0.85

TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-Δ:✓ TV-Δ:✓ TV-Δ:✓ TV-Δ:✗ TV-Δ:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✗ TV-B:✓ TV-B:✓ TV-B:✓ TV-B:✓ TV-C:✓ TV-C:✓ TV-C:✗ TV-C:✗ TV-C:✗ TV-C:✓ TV-C:✗ TV-C:✓

0.6

0.4

TV-Δ:✓ TV-B:✗ TV-C:✗

configuration

TV-Δ:✗ TV-B:✓ TV-C:✓

TV-Δ:✗ TV-B:✓ TV-C:✗

TV-Δ:✓ TV-B:✗ TV-C:✓

TV-Δ:✗ TV-B:✗ TV-C:✓

TV-Δ:✓ TV-B:✓ TV-C:✓

TV-Δ:✗ TV-B:✗ TV-C:✗

TV-Δ:✓ TV-B:✓ TV-C:✗

configuration

(ac) RacketSports (RS)

(ad) StandWalkJump (SWJ)

Figure 9: (Continued.) Visualization of classification accuracy of MambaSL along TI/TV configurations on the 30 UEA datasets, in order of maximum accuracy.

Published as a conference paper at ICLR 2026

C.5

F ULL RESULTS OF M AMBA SL PERFORMANCE ACCORDING TO MODEL DEPTH

Table 8 presents the complete comparison results corresponding to Table 3. Although using three layers yielded the highest average accuracy on the 10 datasets commonly evaluated in TSLib, this observation does not hold when extended to the full set of 30 UEA datasets. When examining the average rank on these 10 datasets, we find that the single-layer configuration consistently outperforms the multi-layer variants, indicating that the higher average accuracy for deeper models is driven by a small subset of datasets. In particular, the HW dataset exhibits a nearly 5%p improvement with 2–3 layers compared to a single layer, which disproportionately affects the average accuracy metric. Table 8: Classification accuracy (%) of MambaSL according to its model depth on 30 datasets from the UEA archives. The best and the second-best are highlighted in bold and underline, respectively. model depth 1 (proposed) 2

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ

41.065 68.785 65.765 80.488 98.649 85.549 92.150 62.778 99.864 93.438 99.333 53.333 100.000 99.652 100.000 70.000 83.206 97.826 95.185 67.000 68.919 64.772 91.111 44.039 67.000 98.889 99.171 31.435 93.421 66.667

41.825 69.098 66.353 80.488 98.378 86.705 91.126 62.222 99.864 93.125 99.333 60.000 100.000 99.721 100.000 68.000 85.496 97.101 94.074 65.000 68.919 65.596 89.444 42.539 66.000 98.333 99.200 30.987 92.763 66.667

avg. acc (10) avg. acc avg. rank (10) avg. rank

78.827 79.819 1.30 1.33

78.853 79.316 1.90 1.80

78.918 79.279 2.10 2.07

26 15 0.035

24 12 0.004

# of win/draw # of lose/draw rank test

Published as a conference paper at ICLR 2026

C.6

C LASSIFICATION RESULTS ON THE UEA BENCHMARK UNDER I NCEPTION T IME SETTING

To address the data-leakage concern raised for TSLib, we trained MambaSL under the InceptionTime protocol, which selects the checkpoint with the lowest training loss (Ismail Fawaz et al., 2020). For comparison, we selected the top three high-performing models from each source—HC2 (Middlehurst et al., 2021), BORF (Spinnato et al., 2024), and our own experiments—to form a representative conventional TSC baseline set. For HC2, we used only the results from fold 0, the official train–test split3 , and for BORF, we approximated the reported accuracies using the number of test samples as described in appendix C.3. Table 9 summarizes the results. Although MambaSL achieved slightly lower average accuracy than HC2 and Hydra, it remained competitive, achieving the largest number of top-1 results (15 datasets) and outperforming every non-DL baseline in direct head-to-head comparisons. These findings indicate that MambaSL maintains strong performance under leakage-free model selection. Since all experimental hyperparameters and optimization settings were fixed to the same values as in our original experiments, the performance differences between Tables 5 and 9 arise solely from the change in the model selection criterion. Table 9: UEA classification accuracy (%) of MambaSL using the InceptionTime protocol (selecting the checkpoint with the lowest training loss). The comparison set includes the top three models from each HC2 (Middlehurst et al., 2021), BORF (Spinnato et al., 2024), and our own evaluations. The best and the second-best are highlighted in bold and underline, respectively. HC2 reported (2021)

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ # of win/draw # of lose/draw rank test acc difference

CIF (2020)

DrCIF (2021)

HC2 (2021)

73.384 62.713 35.647 78.049

69.202 62.003 34.588 79.024

77.186 66.033 54.824 73.171

100.000 100.000 100.000 86.007 87.713 89.078 50.000 49.444 50.000 92.500 90.938 92.813 98.333 98.000 99.333 33.333 33.333 26.667 100.000 100.000 100.000 98.611 44.000 91.603 98.551 98.148 52.000 59.459

98.611 100.000 54.000 56.000 92.366 94.656 97.826 100.000 99.259 98.889 60.000 53.000 52.703 47.297

91.111 57.259 50.000 85.556 96.741 26.543 88.158 40.000

89.444 55.596 44.000 84.444 97.656 30.778 90.132 53.333

93.333 64.274 54.000 89.444 97.913 29.049 90.789 46.667

18 10 0.081 2.058

19 9 0.088 1.416

17 13 0.173 (0.123)

BORF reported (2024) Mini Rocket (2021)

our experiments MR+ Hydra (2023)

MambaSL (proposed)

48.289 63.118 58.935 52.852 54.373 60.076 58.201 59.109 61.890 61.379 60.982 61.294 51.765 36.235 51.412 56.000 56.353 53.529 74.146 79.024 73.171 78.537 76.585 77.073 99.189 97.297 97.838 98.378 97.838 98.378 82.659 100.000 77.457 97.688 98.266 82.659 91.809 87.031 95.222 89.078 86.689 94.881 54.444 51.111 56.667 54.444 58.889 55.000 99.318 97.817 99.318 99.136 99.136 99.045 93.125 93.750 94.063 94.063 94.375 94.063 99.333 99.333 99.333 99.333 99.333 99.333 13.333 40.000 6.667 20.000 26.667 6.667 100.000 100.000 100.000 100.000 100.000 100.000 99.304 98.120 99.234 99.304 99.164 99.373 98.611 98.611 98.611 100.000 100.000 98.611 70.000 48.000 64.000 70.000 68.000 60.000 93.893 93.130 96.183 97.710 98.473 98.473 100.000 100.000 100.000 100.000 100.000 100.000 97.778 99.259 98.889 98.889 98.889 99.259 57.000 47.000 56.000 61.000 55.000 60.000 39.189 51.351 36.486 48.649 51.351 36.486 67.300 66.400 66.700 66.796 67.052 65.744 91.667 89.444 92.778 94.444 93.333 93.333 65.085 56.488 63.585 66.383 66.504 64.274 50.000 50.000 48.000 50.000 54.000 51.000 94.444 82.778 91.667 91.111 90.000 93.333 97.313 96.798 97.656 97.856 97.284 29.407 39.994 35.401 33.015 33.164 33.642 85.526 88.816 90.789 90.132 91.447 90.132 40.000 40.000 60.000 40.000 40.000 66.667

39.163 68.331 59.294 76.585 98.649 83.237 88.737 61.111 99.727 92.500 99.333 46.667 100.000 99.582 100.000 70.000 83.206 97.826 92.593 64.000 62.162 64.524 90.556 25.182 62.000 98.889 99.228 29.824 90.789 60.000

19 13 0.046 2.061

DrCIF (2021)

20 12 0.072 1.777

MR+ Hydra (2023)

20 14 0.082 1.218

HC2 (2021)

18 16 0.307 (0.076)

The full set of reported results is available at timeseriesclassification.com.

Hydra (2023)

19 15 0.240 (0.334)

18 14 0.221 0.470

Published as a conference paper at ICLR 2026

The higher performance observed in our re-evaluation of certain non-DL baselines in Table 9 stems from running three random seeds and reporting the best-performing run. Using only a single seed (seed 0) yields results comparable to those originally reported for HC2 and BORF. To further assess stability, we conducted two additional trials for MambaSL and report the mean and standard deviation across three trials in Table 10. In contrast to Table 9, where HC2 and Hydra showed the strongest performance based on the bestperforming 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 comparisons and achieving the largest number of top-1 results (13 datasets). The Wilcoxon signed-rank test indicates no significant difference between MambaSL and MR+Hydra, suggesting comparable overall effectiveness under this protocol. Table 10: Mean and standard deviation of UEA classification accuracy (%) of MambaSL using the InceptionTime protocol (selecting the checkpoint with the lowest training loss). The comparison set includes the four non-DL models from our own evaluations. The best and the second-best are highlighted in bold and underline, respectively. our experiments

EC FD HW HB JV PEMS SRS1 SRS2 SAD UW AWR AF BM CT CR DDG EW EP ER FM HMD IW LIB LSST MI NATO PD PS RS SWJ # of win/draw # of lose/draw rank test acc difference

ROCKET (2020)

HC2 (2021)

Hydra (2023)

MR+Hydra (2023)

MambaSL (proposed)

99.333 ±0.000 6.667 ±0.000 100.000 ±0.000 99.257 ±0.040 100.000 ±0.000 48.667 ±3.055 90.331 ±1.166 98.792 ±0.418 98.642 ±0.428 42.966 ±1.742 61.076 ±0.512 55.000 ±2.000 49.550 ±3.121 58.863 ±0.378 74.472 ±1.228 38.348 ±0.189 82.883 ±0.780 90.185 ±0.321 63.869 ±0.247 51.667 ±2.082 88.333 ±1.111 83.815 ±0.578 98.132 ±0.100 27.378 ±0.517 90.789 ±0.000 85.324 ±0.341 52.407 ±3.349 99.151 ±0.131 46.667 ±0.000 93.646 ±0.477

99.111 ±0.385 17.778 ±3.849 100.000 ±0.000 98.793 ±0.464 99.537 ±0.802 59.333 ±13.614 94.402 ±4.473 100.000 ±0.000 98.395 ±0.566 50.824 ±3.189 55.789 ±4.842 55.333 ±6.028 36.486 ±10.554 49.451 ±5.831 76.260 ±2.200 64.677 ±3.635 94.324 ±7.022 93.148 ±1.398 62.693 ±3.199 47.667 ±4.041 89.074 ±1.951 85.356 ±10.927 96.884 ±0.722 31.663 ±1.849 87.939 ±2.310 87.827 ±1.097 52.407 ±1.786 98.530 ±0.533 37.778 ±3.849 92.188 ±1.740

99.222 ±0.192 17.778 ±7.698 100.000 ±0.000 99.071 ±0.106 99.537 ±0.802 50.000 ±19.079 93.639 ±4.341 100.000 ±0.000 98.765 ±0.214 50.190 ±3.627 57.454 ±4.565 53.667 ±2.309 38.739 ±15.368 53.098 ±5.335 75.772 ±1.015 61.279 ±5.004 90.270 ±6.642 92.222 ±0.962 64.355 ±3.145 52.000 ±2.000 88.333 ±1.470 91.908 ±10.029 97.484 ±0.619 31.057 ±1.827 90.789 ±0.658 86.121 ±0.710 54.630 ±4.170 98.909 ±0.355 40.000 ±0.000 93.021 ±2.081

99.222 ±0.192 6.667 ±0.000 100.000 ±0.000 99.327 ±0.040 98.611 ±0.000 56.667 ±3.055 97.201 ±1.166 100.000 ±0.000 98.395 ±0.932 58.175 ±2.662 60.036 ±1.122 59.333 ±1.155 34.234 ±2.064 52.588 ±0.941 75.610 ±1.291 65.692 ±0.074 97.928 ±0.563 93.333 ±0.000 63.328 ±1.071 50.333 ±0.577 91.481 ±1.604 80.539 ±2.030 97.170 ±0.103 32.995 ±0.707 89.035 ±1.005 94.084 ±0.859 53.889 ±0.962 98.939 ±0.095 57.778 ±15.396 93.750 ±0.541

99.111 ±0.385 42.222 ±7.698 100.000 ±0.000 99.513 ±0.070 100.000 ±0.000 65.333 ±6.429 82.188 ±1.166 97.343 ±0.837 91.728 ±1.497 37.896 ±1.335 67.263 ±0.928 58.667 ±6.807 63.514 ±2.341 58.667 ±1.087 72.683 ±3.683 64.264 ±0.227 98.198 ±0.780 89.815 ±1.283 18.194 ±6.194 60.667 ±4.163 97.778 ±1.470 82.659 ±3.218 99.123 ±0.259 28.780 ±1.002 89.912 ±1.519 87.372 ±2.365 56.667 ±4.194 99.576 ±0.131 53.333 ±11.547 91.979 ±0.902

18 14 0.127 2.608

18 14 0.169 1.360

18 13 0.185 1.171

16 15 0.461 (0.063)

Published as a conference paper at ICLR 2026

C.7

C LASSIFICATION RESULTS ON ADFTD AND FLAAP

While the UEA archive provides a standardized benchmark for multivariate TSC, it mainly consists of datasets released between 1999 and 2018, many of which are relatively clean and small in scale (Bagnall et al., 2018). To evaluate the robustness of our model beyond these settings, we additionally tested MambaSL on two more recent and domain-specific TSC datasets. Following the evaluation protocol of Medformer (Wang et al., 2024a), we selected ADFTD (Miltiadous et al., 2023) from the medical domain and FLAAP (Kumar & Suresh, 2022) from the human activity recognition (HAR) domain. A summary of both datasets is provided in Table 11. Table 11: Summary of ADFTD and FLAAP datasets Dataset

Domain

Samples

Length

Variables

Classes

File Size

ADFTD (2023) FLAAP (2022)

EEG HAR

69752 13123

256 100

3 10

19 6

2.52GB 60.2MB

Preprocessing and experimental settings were aligned with the official Medformer implementation. We first evaluated MambaSL using the 240 hyperparameter configurations employed for the UEA experiments, after which we conducted four additional trials on some best-performing configurations, resulting in a total of five runs. Table 12 reports the average accuracy and F1-score across these trials. For comparison, we also include the results of the strongly performing baselines reported by Medformer (Wang et al., 2024a): Crossformer (Zhang & Yan, 2023), Reformer (Kitaev et al., 2020), Transformer (Vaswani et al., 2017), TCN (Bai et al., 2018), ModernTCN (Luo & Wang, 2024), and Mamba (Gu & Dao, 2024). Since the original Medformer paper did not report results for TCN, ModernTCN, and Mamba on ADFTD, these entries are left blank. As shown in Table 12, MambaSL performed competitively against previous baselines on both datasets, and achieved the highest accuracy on FLAAP. In particular, we observe more than a 10%p improvement over the vanilla Mamba. Taken together, the results on ADFTD and FLAAP show that MambaSL remains stable across domains and dataset scales, reinforcing that its performance is not limited to the characteristics of the UEA archive. Table 12: Classification accuracy and F1-score (%) of MambaSL and other baselines on ADFTD and FLAAP datasets. The best and the second-best are highlighted in bold and underline, respectively. ADFTD Crossformer (2023) Reformer (2020) Transformer (2017) TCN (2018) ModernTCN (2024) Mamba (2024) Medformer (2024a) MambaSL (proposed)

FLAAP

accuracy

F1-score

accuracy

F1-score

50.45 ±2.31 50.78 ±1.17 50.47 ±2.14

45.50 ±1.70 47.94 ±1.59 48.09 ±1.59

53.27 ±1.54 51.68 ±0.89

50.65 ±1.51 48.73 ±3.05

75.84 ±0.52 71.65 ±1.27 74.96 ±1.25 66.48 ±1.66 74.80 ±0.96 64.87 ±2.78 76.44 ±0.64 77.47 ±1.03

75.52 ±0.66 71.14 ±1.45 74.49 ±1.39 65.29 ±1.74 74.35 ±0.85 64.14 ±2.70 76.25 ±0.65 77.59 ±1.15

Record · ID 19010 · SHA-256 f09ebbe535e97c43
Retrieved via Conceptio — every document is proof-bundled with source, license, and retrieval metadata.