cs.LGFeb 19, 2026

MantisV2: Closing the Zero-Shot Gap in Time Series Classification with Synthetic Data and Test-Time Strategies

Authors: Vasilii Feofanov, Songkang Wen, Jianfeng Zhang, Lujia Pan, Ievgen Redko

Organizations: Huawei Noah’s Ark Lab

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Developing foundation models for time series classification is of high practical relevance, as such models can serve as universal feature extractors for diverse downstream tasks. Although early models such as Mantis have shown the promise of this approach, a substantial performance gap remained between frozen and fine-tuned encoders. In this work, we introduce methods that significantly strengthen zero-shot feature extraction for time series. First, we introduce Mantis+, a variant of Mantis pre-trained entirely on synthetic time series. Second, through controlled ablation studies, we refine the architecture and obtain MantisV2, an improved and more lightweight encoder. Third, we propose an enhanced test-time methodology that leverages intermediate-layer representations and refines output-token aggregation. In addition, we show that performance can be further improved via self-ensembling and cross-model embedding fusion. Extensive experiments on UCR, UEA, Human Activity Recognition (HAR) benchmarks, and EEG datasets show that MantisV2 and Mantis+ consistently outperform prior time series foundation models, achieving state-of-the-art zero-shot performance.

Figures & tables

Mantis+ :https://huggingface.co/paris-noah/MantisPlus
MantisV2 :https://huggingface.co/paris-noah/MantisV2
CauKer 2M :https://huggingface.co/datasets/paris-noah/CauKer2M
Table 1
Figure 1: Final Performance on the UCR Benchmark.
Figure 2: Random Crop Resize.
Pre-training DataNatureSizeUCR Included?UCR acc. (%)
AnomalyReal38KNo0.7473±0.00140.7473_{\pm 0.0014}
Subset of V1 DatasetReal100KNo0.7829±0.00080.7829_{\pm 0.0008}
CauKerSynth100KNo78.81±0.00178.81_{\pm 0.001}
V1 DatasetReal1.89MYes79.21±0.001279.21_{\pm 0.0012}
Table 1: Performance of Mantis on UCR collection with different pre-training datasets.
Figure 3: Ablation study confirming the proposed Token Generator Unit.
Figure 4: Ablation study on different hyperparameters of the architecture.
Figure 5: Comparison of different transformer configurations.
Figure 6: Mantis, layer by layer performance.
Figure 7: Downstream performance evolution: layer by layer, epoch by epoch.
Figure 8: Layer-by-Layer for other models.
Figure 9: Ablation study on output-token aggregation and the summary of all improvements made in this section.
Figure 10: Classification accuracy of different methods in average of 128 univariate UCR datasets.
Figure 11: Classification accuracy of different methods in average of 27 multivariate UEA datasets.
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
Ego4D0.4397±0.0020.4397_{\pm 0.002}NaNNaN0.4068±0.0010.4068_{\pm 0.001}0.5037±0.0010.5037_{\pm 0.001}0.4742±0.00.4742_{\pm 0.0}0.19120.19070.5108±0.00.5108_{\pm 0.0}0.5273 ±0.00.5258±0.0010.5258_{\pm 0.001}
EMOPain0.8826±0.0060.8826_{\pm 0.006}0.78310.78310.8469±0.0020.8469_{\pm 0.002}0.7915±0.0030.7915_{\pm 0.003}0.8075±0.0040.8075_{\pm 0.004}0.83±0.0020.83_{\pm 0.002}0.8225±0.0030.8225_{\pm 0.003}0.8901±0.0050.8901_{\pm 0.005}0.8939 ±0.0020.8798±0.0090.8798_{\pm 0.009}
HHAR-ID0.9738±0.0010.9738_{\pm 0.001}0.89380.90730.9299±0.00.9299_{\pm 0.0}0.9481±0.00.9481_{\pm 0.0}0.9475±0.0020.9475_{\pm 0.002}0.9338±0.0010.9338_{\pm 0.001}0.9461±0.0010.9461_{\pm 0.001}0.9808±0.0010.9808_{\pm 0.001}0.9835±0.0010.9835_{\pm 0.001}0.9845 ±0.001
HHAR-OOD0.4808±0.0080.4808_{\pm 0.008}0.53110.54410.3081±0.0010.3081_{\pm 0.001}0.5135±0.0140.5135_{\pm 0.014}0.4174±0.0080.4174_{\pm 0.008}0.3748±0.0070.3748_{\pm 0.007}0.4031±0.010.4031_{\pm 0.01}0.56±0.0050.56_{\pm 0.005}0.5624±0.0160.5624_{\pm 0.016}0.5822 ±0.005
MP80.572±0.0080.572_{\pm 0.008}0.62350.61850.6392±0.0110.6392_{\pm 0.011}0.5994±0.0160.5994_{\pm 0.016}0.5966±0.0060.5966_{\pm 0.006}0.6022±0.0090.6022_{\pm 0.009}0.5966±0.0030.5966_{\pm 0.003}0.6504±0.0080.6504_{\pm 0.008}0.6403±0.0060.6403_{\pm 0.006}0.6857 ±0.012
MP500.3602±0.0160.3602_{\pm 0.016}0.56640.50420.7434 ±0.0170.6644±0.010.6644_{\pm 0.01}0.6639±0.0070.6639_{\pm 0.007}0.6908±0.0080.6908_{\pm 0.008}0.6801±0.0080.6801_{\pm 0.008}0.6913±0.0120.6913_{\pm 0.012}0.7227±0.0080.7227_{\pm 0.008}0.7345±0.0070.7345_{\pm 0.007}
Table 2: Classification accuracy of different methods on a collection of HAR datasets. Boldface highlights the best method per row while NaN indicates that the model have not passed the scale.
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
Blink0.9963±0.0010.9963_{\pm 0.001}0.91780.89780.9674±0.0030.9674_{\pm 0.003}0.997 ±0.0010.9733±0.010.9733_{\pm 0.01}0.9852±0.0060.9852_{\pm 0.006}0.98±0.0020.98_{\pm 0.002}0.66±0.0060.66_{\pm 0.006}0.9778±0.0020.9778_{\pm 0.002}0.9956±0.0020.9956_{\pm 0.002}
CAP-ID0.751±0.0010.751_{\pm 0.001}NaNNaN0.7374±0.0020.7374_{\pm 0.002}0.8104±0.0020.8104_{\pm 0.002}0.8179±0.0010.8179_{\pm 0.001}0.7983±0.0010.7983_{\pm 0.001}0.8126±0.0010.8126_{\pm 0.001}0.8044±0.00.8044_{\pm 0.0}0.8189 ±0.0010.8152±0.0010.8152_{\pm 0.001}
CAP-OOD0.71±0.0030.71_{\pm 0.003}NaNNaN0.7408±0.0010.7408_{\pm 0.001}0.7821±0.0010.7821_{\pm 0.001}0.7857±0.0010.7857_{\pm 0.001}0.7781±0.0010.7781_{\pm 0.001}0.784±0.0010.784_{\pm 0.001}0.767±0.0020.767_{\pm 0.002}0.791 ±0.0010.7859±0.00.7859_{\pm 0.0}
Epilepsy-EEG0.9507±0.0020.9507_{\pm 0.002}0.94960.94470.952±0.0010.952_{\pm 0.001}0.9614 ±0.0010.95±0.0010.95_{\pm 0.001}0.9548±0.00.9548_{\pm 0.0}0.9495±0.0010.9495_{\pm 0.001}0.9234±0.0020.9234_{\pm 0.002}0.9518±0.0010.9518_{\pm 0.001}0.9558±0.0010.9558_{\pm 0.001}
FingerMovements0.4967±0.0640.4967_{\pm 0.064}0.50.490.53±0.020.53_{\pm 0.02}0.52±0.040.52_{\pm 0.04}0.5233±0.0210.5233_{\pm 0.021}0.5367±0.0450.5367_{\pm 0.045}0.5333±0.0590.5333_{\pm 0.059}0.5267±0.0150.5267_{\pm 0.015}0.51±0.0260.51_{\pm 0.026}0.55 ±0.01
PCL-ID0.5241±0.0030.5241_{\pm 0.003}NaNNaN0.5431±0.0120.5431_{\pm 0.012}0.5272±0.0110.5272_{\pm 0.011}0.5308±0.0020.5308_{\pm 0.002}0.5347±0.0070.5347_{\pm 0.007}0.5385±0.0050.5385_{\pm 0.005}0.5615±0.0030.5615_{\pm 0.003}0.5764 ±0.0060.5716±0.0030.5716_{\pm 0.003}
Table 3: Classification accuracy of different methods on a collection of EEG datasets. Boldface highlights the best method per row while NaN indicates that the model have not passed the scale.
# of Params.TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
Original7.2M27.1M341.2M35.3M119.5M630.8M843.4M2.4M8.1M4.2M
After Pruning--161.4M16.5M44M276.6M180.3M2M2.9M2.2M
Table 4: Comparison of Model Sizes.
Figure 12: The inference time of the considered models as a function of the number of samples.
Figure 13: Motivation to look at different interpolation sizes that, in the context of our architecture, imply different degree of overlap between tokens.
Figure 14: Self-Ensembling Ablation Study.
Figure 15: Performance comparison on UCR.
Figure 16: Comparison with more baselines (taken from Goswami et al., 2024 ) on 91 UCR datasets.
Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset# of ChannelsSeq. LengthTrain SizeTest SizeNum. classes
Ego4D610002470956699431
EMOPain302009683553
HHAR-ID6500871634196
HHAR-OOD65001115021546
MP8816114265954
MP505016114265954
Appendix
Table 5: Characteristics of HAR datasets.
Dataset# of ChannelsSeq. LengthTrain SizeTest SizeNum. Classes
Blink45105004502
CAP-ID19300025748100986
CAP-OOD1930002739382656
Epilepsy-EEG117860114202
FingerMovements28503161002
PCL-ID487501440556502
Appendix
Table 6: Characteristics of EEG datasets.
Figure 17: Catch22+ ablation study.
Figure 18: Layer-by-layer downstream performance (UCR) over 1000 epochs with 100K pre-training samples.
Figure 19: Final pre-training. 200 epochs with 2M pre-training samples.
Figure 20: Ablation study on output-token aggregation when the last transformer layer is used.
LayerV1 TransformerV2 Transformer
Number
10.78720.7669
20.79790.7987
30.79890.7994
40.79270.7986
50.79270.7919
Appendix
Table 7: Performance comparison between the first and the second version of transformer architecture (keeping all the other parameters of Mantis architecture fixed as in the first version). Both architectures were pre-trained on 1 million synthetic examples generated by CauKer, while the last epoch accuracy on UCR benchmark is reported.
Figure 21: Fine-tuning results for different configurations of Mantis+ and MantisV2. "Truncated Model" means that all transformer layers that succeed the intermediate layer with the highest zero-shot performance are removed. "Full Model" refers to keeping all the layers.
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
ACSF10.8233±0.0210.8233_{\pm 0.021}0.80.810.82±0.0260.82_{\pm 0.026}0.85±0.010.85_{\pm 0.01}0.8467±0.0120.8467_{\pm 0.012}0.85±0.00.85_{\pm 0.0}0.8667 ±0.0060.75±0.020.75_{\pm 0.02}0.8033±0.0060.8033_{\pm 0.006}0.8233±0.0120.8233_{\pm 0.012}
Adiac0.734±0.0070.734_{\pm 0.007}0.80310.80310.7894±0.0030.7894_{\pm 0.003}0.7852±0.0110.7852_{\pm 0.011}0.8107±0.0090.8107_{\pm 0.009}0.7059±0.0070.7059_{\pm 0.007}0.6547±0.0180.6547_{\pm 0.018}0.7357±0.0060.7357_{\pm 0.006}0.8321±0.0060.8321_{\pm 0.006}0.8483 ±0.003
AllGestureWiimoteX0.5957±0.010.5957_{\pm 0.01}0.62290.50430.6138±0.0020.6138_{\pm 0.002}0.6495±0.0120.6495_{\pm 0.012}0.6038±0.0050.6038_{\pm 0.005}0.6029±0.0110.6029_{\pm 0.011}0.6676±0.0020.6676_{\pm 0.002}0.6505±0.0050.6505_{\pm 0.005}0.6414±0.0120.6414_{\pm 0.012}0.6843 ±0.006
AllGestureWiimoteY0.6319±0.0160.6319_{\pm 0.016}0.63290.51140.6624±0.0040.6624_{\pm 0.004}0.7052 ±0.0190.6419±0.0020.6419_{\pm 0.002}0.6514±0.010.6514_{\pm 0.01}0.6805±0.0110.6805_{\pm 0.011}0.63±0.010.63_{\pm 0.01}0.691±0.0060.691_{\pm 0.006}0.691±0.0040.691_{\pm 0.004}
AllGestureWiimoteZ0.5462±0.0050.5462_{\pm 0.005}0.53290.45290.5767±0.0070.5767_{\pm 0.007}0.6057±0.0010.6057_{\pm 0.001}0.6205±0.0040.6205_{\pm 0.004}0.6129±0.0140.6129_{\pm 0.014}0.6067±0.0110.6067_{\pm 0.011}0.6224±0.0090.6224_{\pm 0.009}0.6452±0.0040.6452_{\pm 0.004}0.6881 ±0.001
ArrowHead0.741±0.0120.741_{\pm 0.012}0.75430.74290.7943±0.0150.7943_{\pm 0.015}0.7886±0.010.7886_{\pm 0.01}0.779±0.0370.779_{\pm 0.037}0.7638±0.0120.7638_{\pm 0.012}0.7714±0.0060.7714_{\pm 0.006}0.7733±0.0230.7733_{\pm 0.023}0.7714±0.0060.7714_{\pm 0.006}0.8057 ±0.015
Appendix
Table 8: Zero-shot feature extraction results on UCR (Random Forest). First Part.
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
ItalyPowerDemand0.9537±0.0020.9537_{\pm 0.002}0.96990.96310.9482±0.0040.9482_{\pm 0.004}0.9624±0.0040.9624_{\pm 0.004}0.9559±0.0030.9559_{\pm 0.003}0.8776±0.0030.8776_{\pm 0.003}0.9248±0.0020.9248_{\pm 0.002}0.8737±0.0040.8737_{\pm 0.004}0.9002±0.0050.9002_{\pm 0.005}0.9248±0.0040.9248_{\pm 0.004}
LargeKitchenAppliances0.8169±0.0090.8169_{\pm 0.009}0.63730.6960.7342±0.0070.7342_{\pm 0.007}0.8053±0.0050.8053_{\pm 0.005}0.8729 ±0.0030.816±0.0030.816_{\pm 0.003}0.8702±0.0020.8702_{\pm 0.002}0.7556±0.0090.7556_{\pm 0.009}0.8133±0.0110.8133_{\pm 0.011}0.7662±0.0060.7662_{\pm 0.006}
Lightning20.7322±0.0090.7322_{\pm 0.009}0.67210.70490.7869±0.0330.7869_{\pm 0.033}0.7432±0.0090.7432_{\pm 0.009}0.7268±0.0090.7268_{\pm 0.009}0.7596±0.0090.7596_{\pm 0.009}0.7432±0.0250.7432_{\pm 0.025}0.6776±0.0090.6776_{\pm 0.009}0.7923 ±0.0090.7268±0.0090.7268_{\pm 0.009}
Lightning70.7671 ±0.0140.69860.73970.7352±0.0290.7352_{\pm 0.029}0.7352±0.0160.7352_{\pm 0.016}0.653±0.0210.653_{\pm 0.021}0.7489±0.0290.7489_{\pm 0.029}0.7306±0.0160.7306_{\pm 0.016}0.6575±0.0270.6575_{\pm 0.027}0.7397±0.0270.7397_{\pm 0.027}0.7215±0.0440.7215_{\pm 0.044}
Mallat0.9555±0.0090.9555_{\pm 0.009}0.96890.94840.9164±0.0160.9164_{\pm 0.016}0.9166±0.0270.9166_{\pm 0.027}0.9016±0.0090.9016_{\pm 0.009}0.9032±0.0070.9032_{\pm 0.007}0.8387±0.0050.8387_{\pm 0.005}0.8311±0.0050.8311_{\pm 0.005}0.8708±0.010.8708_{\pm 0.01}0.8887±0.0020.8887_{\pm 0.002}
Meat0.9222±0.010.9222_{\pm 0.01}0.98330.93330.9222±0.010.9222_{\pm 0.01}0.8611±0.010.8611_{\pm 0.01}0.9333±0.0170.9333_{\pm 0.017}0.8389±0.0380.8389_{\pm 0.038}0.8667±0.00.8667_{\pm 0.0}0.9278±0.0190.9278_{\pm 0.019}0.8944±0.010.8944_{\pm 0.01}0.9111±0.0250.9111_{\pm 0.025}
Appendix
Table 9: Zero-shot feature extraction results on UCR (Random Forest). Second Part.
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
ArticularyWordRecognition0.96±0.0070.96_{\pm 0.007}0.930.910.9844±0.0020.9844_{\pm 0.002}0.99±0.00.99_{\pm 0.0}0.9933±0.0030.9933_{\pm 0.003}0.9833±0.0030.9833_{\pm 0.003}0.98±0.0030.98_{\pm 0.003}0.9911±0.0040.9911_{\pm 0.004}0.9956 ±0.0020.9922±0.0020.9922_{\pm 0.002}
BasicMotions1.0 ±0.01.00.950.975±0.00.975_{\pm 0.0}0.9667±0.0140.9667_{\pm 0.014}1.0 ±0.00.9833±0.0140.9833_{\pm 0.014}1.0 ±0.01.0 ±0.01.0 ±0.01.0 ±0.0
CharacterTrajectories0.9814 ±0.0010.96730.96940.9698±0.0020.9698_{\pm 0.002}0.9603±0.00.9603_{\pm 0.0}0.9631±0.0010.9631_{\pm 0.001}0.9373±0.0020.9373_{\pm 0.002}0.9517±0.0020.9517_{\pm 0.002}0.967±0.0020.967_{\pm 0.002}0.9761±0.0010.9761_{\pm 0.001}0.9761±0.00.9761_{\pm 0.0}
Cricket0.9444±0.0140.9444_{\pm 0.014}0.81940.84720.9722±0.00.9722_{\pm 0.0}0.9954±0.0080.9954_{\pm 0.008}0.9861±0.00.9861_{\pm 0.0}1.0 ±0.00.9954±0.0080.9954_{\pm 0.008}0.9907±0.0080.9907_{\pm 0.008}0.9861±0.00.9861_{\pm 0.0}0.9861±0.00.9861_{\pm 0.0}
DuckDuckGeese0.5133±0.0230.5133_{\pm 0.023}0.320.280.4467±0.0420.4467_{\pm 0.042}0.42±0.020.42_{\pm 0.02}0.4133±0.0230.4133_{\pm 0.023}0.4333±0.0230.4333_{\pm 0.023}0.44±0.040.44_{\pm 0.04}0.4267±0.0230.4267_{\pm 0.023}0.5067±0.0310.5067_{\pm 0.031}0.54 ±0.02
ERing0.8827±0.0220.8827_{\pm 0.022}0.88890.85190.9679±0.0080.9679_{\pm 0.008}0.9765±0.0080.9765_{\pm 0.008}0.9877 ±0.0040.9815±0.0040.9815_{\pm 0.004}0.963±0.0070.963_{\pm 0.007}0.9765±0.0040.9765_{\pm 0.004}0.9877 ±0.0060.9877 ±0.002
Appendix
Table 10: Zero-shot feature extraction results on UEA-27 (Random Forest).
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
ACSF10.70.80.810.750.750.860.830.830.70.790.76
Adiac0.6650.80310.80310.77750.78260.82860.71870.70590.80050.84140.8363
AllGestureWiimoteX0.54860.62290.50430.66140.70140.68430.670.69430.60710.71710.73
AllGestureWiimoteY0.52290.63290.51140.70290.73140.70430.73710.72290.65290.74140.7471
AllGestureWiimoteZ0.48290.53290.45290.60570.63430.660.67710.65860.54140.66570.6814
ArrowHead0.71430.75430.74290.77710.78290.85140.82290.82290.72570.84570.8514
Appendix
Table 11: Zero-shot feature extraction results on UCR (Logistic Regression). First Part.
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
ItalyPowerDemand0.93880.96990.96310.95430.96020.9640.94070.93780.91160.92810.9456
LargeKitchenAppliances0.83730.63730.6960.760.82930.86130.8720.89330.77070.8240.816
Lightning20.75410.67210.70490.78690.73770.78690.81970.83610.73770.77050.7869
Lightning70.69860.69860.73970.73970.80820.79450.84930.76710.78080.83560.863
Mallat0.88660.96890.94840.91860.95480.94670.97650.96420.84860.90490.9198
Meat0.93330.98330.93331.00.90.93330.81670.850.93330.98330.9667
Appendix
Table 12: Zero-shot feature extraction results on UCR (Logistic Regression). Second Part.
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
ArticularyWordRecognition0.93330.930.910.98670.990.99330.98330.97670.99330.99330.9933
BasicMotions1.01.00.951.01.01.01.01.01.01.01.0
CharacterTrajectories0.9840.96730.96940.9840.96940.97420.96590.97140.98470.98120.9868
Cricket0.90280.81940.84720.98611.01.01.01.01.00.98610.9861
DuckDuckGeese0.420.320.280.480.380.340.480.560.50.540.5
ERing0.90740.88890.85190.97040.97780.97780.98520.98150.98150.99260.9889
Appendix
Table 13: Zero-shot feature extraction results on UEA-27 (Logistic Regression).
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
Ego4D0.3719±0.00.3719_{\pm 0.0}NaNNaN0.4324±0.00.4324_{\pm 0.0}0.5267±0.00.5267_{\pm 0.0}0.5216±0.00.5216_{\pm 0.0}0.13940.12880.5248±0.00.5248_{\pm 0.0}0.5519±0.00.5519_{\pm 0.0}0.5551 ±0.0
EMOPain0.7915±0.00.7915_{\pm 0.0}0.78310.78310.8225±0.00.8225_{\pm 0.0}0.8254±0.00.8254_{\pm 0.0}0.8338±0.00.8338_{\pm 0.0}0.8507 ±0.00.8028±0.00.8028_{\pm 0.0}0.7831±0.00.7831_{\pm 0.0}0.8451±0.00.8451_{\pm 0.0}0.8169±0.00.8169_{\pm 0.0}
HHAR-ID0.9111±0.00.9111_{\pm 0.0}0.89380.90730.9491±0.00.9491_{\pm 0.0}0.9708±0.00.9708_{\pm 0.0}0.9699±0.00.9699_{\pm 0.0}0.9661±0.00.9661_{\pm 0.0}0.9693±0.00.9693_{\pm 0.0}0.9751±0.00.9751_{\pm 0.0}0.9871±0.00.9871_{\pm 0.0}0.9883 ±0.0
HHAR-OOD0.4085±0.00.4085_{\pm 0.0}0.53110.54410.3329±0.00.3329_{\pm 0.0}0.3338±0.00.3338_{\pm 0.0}0.3691±0.00.3691_{\pm 0.0}0.318±0.00.318_{\pm 0.0}0.3222±0.00.3222_{\pm 0.0}0.5084±0.00.5084_{\pm 0.0}0.4907±0.00.4907_{\pm 0.0}0.5604 ±0.0
MP80.642±0.00.642_{\pm 0.0}0.62350.61850.595±0.00.595_{\pm 0.0}0.6134±0.00.6134_{\pm 0.0}0.6387±0.00.6387_{\pm 0.0}0.6403±0.00.6403_{\pm 0.0}0.642±0.00.642_{\pm 0.0}0.6084±0.00.6084_{\pm 0.0}0.6588 ±0.00.6521±0.00.6521_{\pm 0.0}
MP500.4471±0.00.4471_{\pm 0.0}0.56640.50420.7294±0.00.7294_{\pm 0.0}0.7563 ±0.00.7345±0.00.7345_{\pm 0.0}0.7193±0.00.7193_{\pm 0.0}0.6857±0.00.6857_{\pm 0.0}0.6992±0.00.6992_{\pm 0.0}0.6924±0.00.6924_{\pm 0.0}0.716±0.00.716_{\pm 0.0}
Appendix
Table 14: Zero-shot feature extraction results on HAR (Logistic Regression).
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2
Blink0.960.91780.89780.98670.98891.00.98890.98220.59110.94890.9733
CAP0.6415NaNNaN0.73740.8010.81680.81390.8150.77670.80780.7945
CAP-OOD0.6704NaNNaN0.69460.72240.75350.72230.7290.6330.64150.6391
Epilepsy-EEG0.9470.94960.94470.93660.95630.94140.93220.91860.9070.94240.9312
FingerMovements0.520.50.490.570.620.60.570.530.610.570.64
PCL0.5402NaNNaN0.56160.56670.55960.56850.57310.57540.5750.5931
Appendix
Table 15: Zero-shot feature extraction results on EEG (Logistic Regression).
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2SE-Mantis+SE-MantisV2MantisV2 &MantisV2 &MantisV2-FT
              TiViT-HTiConvNext 
ACSF10.8233±0.0210.8233_{\pm 0.021}0.80.810.750.750.860.830.830.70.790.760.80.770.810.820.77±0.010.77_{\pm 0.01}
Adiac0.734±0.0070.734_{\pm 0.007}0.80310.80310.77750.78260.82860.71870.70590.80050.84140.83630.83120.84650.75190.77490.8321±0.0070.8321_{\pm 0.007}
AllGestureWiimoteX0.5957±0.010.5957_{\pm 0.01}0.62290.50430.66140.70140.68430.670.69430.60710.71710.730.690.68710.72290.73570.779 ±0.006
AllGestureWiimoteY0.6319±0.0160.6319_{\pm 0.016}0.63290.51140.70290.73140.70430.73710.72290.65290.74140.74710.730.72710.75570.74860.7905 ±0.004
AllGestureWiimoteZ0.5462±0.0050.5462_{\pm 0.005}0.53290.45290.60570.63430.660.67710.65860.54140.66570.68140.63290.63570.68860.70430.7505 ±0.017
Appendix
Table 16: The final comparison on UCR. First Part.
 Catch22+TabPFNTabICLMOMENTTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2SE-Mantis+SE-MantisV2MantisV2 &MantisV2 &MantisV2-FT
              TiViT-HTiConvNext 
ItalyPowerDemand0.9537±0.0020.9537_{\pm 0.002}0.96990.96310.95430.96020.9640.94070.93780.91160.92810.94560.94360.94750.94270.94460.9517±0.0010.9517_{\pm 0.001}
LargeKitchenAppliances0.8169±0.0090.8169_{\pm 0.009}0.63730.6960.760.82930.86130.8720.89330.77070.8240.8160.8320.8160.85070.8880.8596±0.0080.8596_{\pm 0.008}
Lightning20.7322±0.0090.7322_{\pm 0.009}0.67210.70490.78690.73770.78690.81970.83610.73770.77050.78690.83610.77050.85250.80330.847±0.0250.847_{\pm 0.025}
Lightning70.7671±0.0140.7671_{\pm 0.014}0.69860.73970.73970.80820.79450.84930.76710.78080.83560.8630.8630.82190.87670.79450.7717±0.0320.7717_{\pm 0.032}
Mallat0.9555±0.0090.9555_{\pm 0.009}0.96890.94840.91860.95480.94670.97650.96420.84860.90490.91980.93390.96630.9740.96970.9512±0.0210.9512_{\pm 0.021}
Appendix
Table 17: The final comparison on UCR. Second Part.
 TS2VecT-LossTNCTS-TCCCNNEncoderFCNMLPResNetTWIESNDTWOfficialImprovedCatch22+TabPFNTabICLTiRexChronos2TiViT-HTiConvNextNuTimeMantis+MantisV2SE-Mantis+SE-MantisV2MantisV2 &MantisV2 &
            MOMENTMOMENT            TiViT-HTiConvNext
Adiac0.7620.6750.7670.7670.3930.3180.8410.3910.8330.4280.6040.6880.77750.7340.80310.80310.78260.82860.71870.70590.80050.84140.83630.83120.84650.75190.7749
AllGestureWiimoteX0.7770.7630.6970.6970.4110.4750.7130.4770.7410.5220.7160.6070.66140.59570.62290.50430.70140.68430.670.69430.60710.71710.730.690.68710.72290.7357
AllGestureWiimoteY0.7930.7260.7410.7410.4790.5090.7840.5710.7940.60.7290.6660.70290.63190.63290.51140.73140.70430.73710.72290.65290.74140.74710.730.72710.75570.7486
AllGestureWiimoteZ0.7460.7230.6890.6890.3750.3960.6920.4390.7260.5160.6430.5370.60570.54620.53290.45290.63430.660.67710.65860.54140.66570.68140.63290.63570.68860.7043
ArrowHead0.8570.7660.7370.7370.7170.630.8430.7840.8380.6890.7030.7430.77710.7410.75430.74290.78290.85140.82290.82290.72570.84570.85140.840.85140.880.8457
Appendix
Table 18: Comparison with more baselines (taken from Goswami et al., 2024 ) on 91 UCR datasets.
More in this topic

Time Series Classification

Time series classification assigns labels to sequential data, using temporal patterns to support decisions in health, sensing, science, and industry.

See all 170 papers

Explore similar work

CardsList

  1. Mantis: Lightweight Foundation Model for Time Series Classification

    Feb 21, 2025Vasilii Feofanov, Songkang Wen, Shifeng Xie +10Time Series ClassificationSelf-Supervised Time Series Representation Learning

    1. Huawei Noah’s Ark Lab · 3. Paris Cit´e University · 4. Helmholtz Munich +3

  2. UTICA: Multi-Objective Self-Distllation Foundation Model Pretraining for Time Series Classification

    Mar 2, 2026Yessin Moakher, Youssef Attia El Hili, Vasilii FeofanovSelf-Supervised Time Series Representation LearningTime-Series Foundation Models

    Ecole Polytechnique, France · Huawei Noah’s Ark Lab

  3. Rethinking Zero-Shot Time Series Classification: From Task-specific Classifiers to In-Context Inference

    Jan 31, 2026Juntao Fang, Shifeng Xie, Shengbin Nie +7Time Series ClassificationIn-Context Learning

    Guangdong University of Technology, Guangzhou, China · Huawei Noah’s Ark Lab, Paris, France · Paris Descartes University, Paris, France +2