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

Vasilii Feofanov; Songkang Wen; Jianfeng Zhang; Lujia Pan; Ievgen Redko

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

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.

Pre-training Data	Nature	Size	UCR Included?	UCR acc. (%)
Anomaly	Real	38K	No	$0.7473_{\pm 0.0014}$
Subset of V1 Dataset	Real	100K	No	$0.7829_{\pm 0.0008}$
CauKer	Synth	100K	No	$78.81_{\pm 0.001}$
V1 Dataset	Real	1.89M	Yes	$79.21_{\pm 0.0012}$

Table 1: Performance of Mantis on UCR collection with different pre-training datasets.

Figure 3: Ablation study confirming the proposed Token Generator Unit.

Panel of Figure 4 — 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.

Panel of Figure 9 — 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+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
Ego4D	$0.4397_{\pm 0.002}$	NaN	NaN	$0.4068_{\pm 0.001}$	$0.5037_{\pm 0.001}$	$0.4742_{\pm 0.0}$	0.1912	0.1907	$0.5108_{\pm 0.0}$	0.5273 ±0.0	$0.5258_{\pm 0.001}$
EMOPain	$0.8826_{\pm 0.006}$	0.7831	0.7831	$0.8469_{\pm 0.002}$	$0.7915_{\pm 0.003}$	$0.8075_{\pm 0.004}$	$0.83_{\pm 0.002}$	$0.8225_{\pm 0.003}$	$0.8901_{\pm 0.005}$	0.8939 ±0.002	$0.8798_{\pm 0.009}$
HHAR-ID	$0.9738_{\pm 0.001}$	0.8938	0.9073	$0.9299_{\pm 0.0}$	$0.9481_{\pm 0.0}$	$0.9475_{\pm 0.002}$	$0.9338_{\pm 0.001}$	$0.9461_{\pm 0.001}$	$0.9808_{\pm 0.001}$	$0.9835_{\pm 0.001}$	0.9845 ±0.001
HHAR-OOD	$0.4808_{\pm 0.008}$	0.5311	0.5441	$0.3081_{\pm 0.001}$	$0.5135_{\pm 0.014}$	$0.4174_{\pm 0.008}$	$0.3748_{\pm 0.007}$	$0.4031_{\pm 0.01}$	$0.56_{\pm 0.005}$	$0.5624_{\pm 0.016}$	0.5822 ±0.005
MP8	$0.572_{\pm 0.008}$	0.6235	0.6185	$0.6392_{\pm 0.011}$	$0.5994_{\pm 0.016}$	$0.5966_{\pm 0.006}$	$0.6022_{\pm 0.009}$	$0.5966_{\pm 0.003}$	$0.6504_{\pm 0.008}$	$0.6403_{\pm 0.006}$	0.6857 ±0.012
MP50	$0.3602_{\pm 0.016}$	0.5664	0.5042	0.7434 ±0.017	$0.6644_{\pm 0.01}$	$0.6639_{\pm 0.007}$	$0.6908_{\pm 0.008}$	$0.6801_{\pm 0.008}$	$0.6913_{\pm 0.012}$	$0.7227_{\pm 0.008}$	$0.7345_{\pm 0.007}$

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+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
Blink	$0.9963_{\pm 0.001}$	0.9178	0.8978	$0.9674_{\pm 0.003}$	0.997 ±0.001	$0.9733_{\pm 0.01}$	$0.9852_{\pm 0.006}$	$0.98_{\pm 0.002}$	$0.66_{\pm 0.006}$	$0.9778_{\pm 0.002}$	$0.9956_{\pm 0.002}$
CAP-ID	$0.751_{\pm 0.001}$	NaN	NaN	$0.7374_{\pm 0.002}$	$0.8104_{\pm 0.002}$	$0.8179_{\pm 0.001}$	$0.7983_{\pm 0.001}$	$0.8126_{\pm 0.001}$	$0.8044_{\pm 0.0}$	0.8189 ±0.001	$0.8152_{\pm 0.001}$
CAP-OOD	$0.71_{\pm 0.003}$	NaN	NaN	$0.7408_{\pm 0.001}$	$0.7821_{\pm 0.001}$	$0.7857_{\pm 0.001}$	$0.7781_{\pm 0.001}$	$0.784_{\pm 0.001}$	$0.767_{\pm 0.002}$	0.791 ±0.001	$0.7859_{\pm 0.0}$
Epilepsy-EEG	$0.9507_{\pm 0.002}$	0.9496	0.9447	$0.952_{\pm 0.001}$	0.9614 ±0.001	$0.95_{\pm 0.001}$	$0.9548_{\pm 0.0}$	$0.9495_{\pm 0.001}$	$0.9234_{\pm 0.002}$	$0.9518_{\pm 0.001}$	$0.9558_{\pm 0.001}$
FingerMovements	$0.4967_{\pm 0.064}$	0.5	0.49	$0.53_{\pm 0.02}$	$0.52_{\pm 0.04}$	$0.5233_{\pm 0.021}$	$0.5367_{\pm 0.045}$	$0.5333_{\pm 0.059}$	$0.5267_{\pm 0.015}$	$0.51_{\pm 0.026}$	0.55 ±0.01
PCL-ID	$0.5241_{\pm 0.003}$	NaN	NaN	$0.5431_{\pm 0.012}$	$0.5272_{\pm 0.011}$	$0.5308_{\pm 0.002}$	$0.5347_{\pm 0.007}$	$0.5385_{\pm 0.005}$	$0.5615_{\pm 0.003}$	0.5764 ±0.006	$0.5716_{\pm 0.003}$

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.	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
Original	7.2M	27.1M	341.2M	35.3M	119.5M	630.8M	843.4M	2.4M	8.1M	4.2M
After Pruning	-	-	161.4M	16.5M	44M	276.6M	180.3M	2M	2.9M	2.2M

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.

Panel of Figure 15 — 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 Channels	Seq. Length	Train Size	Test Size	Num. classes
Ego4D	6	1000	247095	66994	31
EMOPain	30	200	968	355	3
HHAR-ID	6	500	8716	3419	6
HHAR-OOD	6	500	11150	2154	6
MP8	8	161	1426	595	4
MP50	50	161	1426	595	4

Appendix

Table 5: Characteristics of HAR datasets.

Dataset	# of Channels	Seq. Length	Train Size	Test Size	Num. Classes
Blink	4	510	500	450	2
CAP-ID	19	3000	25748	10098	6
CAP-OOD	19	3000	27393	8265	6
Epilepsy-EEG	1	178	60	11420	2
FingerMovements	28	50	316	100	2
PCL-ID	48	750	14405	5650	2

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.

Layer	V1 Transformer	V2 Transformer
Number	V1 Transformer	V2 Transformer
1	0.7872	0.7669
2	0.7979	0.7987
3	0.7989	0.7994
4	0.7927	0.7986
5	0.7927	0.7919

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.

Panel of Figure 21 — 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+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
ACSF1	$0.8233_{\pm 0.021}$	0.8	0.81	$0.82_{\pm 0.026}$	$0.85_{\pm 0.01}$	$0.8467_{\pm 0.012}$	$0.85_{\pm 0.0}$	0.8667 ±0.006	$0.75_{\pm 0.02}$	$0.8033_{\pm 0.006}$	$0.8233_{\pm 0.012}$
Adiac	$0.734_{\pm 0.007}$	0.8031	0.8031	$0.7894_{\pm 0.003}$	$0.7852_{\pm 0.011}$	$0.8107_{\pm 0.009}$	$0.7059_{\pm 0.007}$	$0.6547_{\pm 0.018}$	$0.7357_{\pm 0.006}$	$0.8321_{\pm 0.006}$	0.8483 ±0.003
AllGestureWiimoteX	$0.5957_{\pm 0.01}$	0.6229	0.5043	$0.6138_{\pm 0.002}$	$0.6495_{\pm 0.012}$	$0.6038_{\pm 0.005}$	$0.6029_{\pm 0.011}$	$0.6676_{\pm 0.002}$	$0.6505_{\pm 0.005}$	$0.6414_{\pm 0.012}$	0.6843 ±0.006
AllGestureWiimoteY	$0.6319_{\pm 0.016}$	0.6329	0.5114	$0.6624_{\pm 0.004}$	0.7052 ±0.019	$0.6419_{\pm 0.002}$	$0.6514_{\pm 0.01}$	$0.6805_{\pm 0.011}$	$0.63_{\pm 0.01}$	$0.691_{\pm 0.006}$	$0.691_{\pm 0.004}$
AllGestureWiimoteZ	$0.5462_{\pm 0.005}$	0.5329	0.4529	$0.5767_{\pm 0.007}$	$0.6057_{\pm 0.001}$	$0.6205_{\pm 0.004}$	$0.6129_{\pm 0.014}$	$0.6067_{\pm 0.011}$	$0.6224_{\pm 0.009}$	$0.6452_{\pm 0.004}$	0.6881 ±0.001
ArrowHead	$0.741_{\pm 0.012}$	0.7543	0.7429	$0.7943_{\pm 0.015}$	$0.7886_{\pm 0.01}$	$0.779_{\pm 0.037}$	$0.7638_{\pm 0.012}$	$0.7714_{\pm 0.006}$	$0.7733_{\pm 0.023}$	$0.7714_{\pm 0.006}$	0.8057 ±0.015

Appendix

Table 8: Zero-shot feature extraction results on UCR (Random Forest). First Part.

	Catch22+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
ItalyPowerDemand	$0.9537_{\pm 0.002}$	0.9699	0.9631	$0.9482_{\pm 0.004}$	$0.9624_{\pm 0.004}$	$0.9559_{\pm 0.003}$	$0.8776_{\pm 0.003}$	$0.9248_{\pm 0.002}$	$0.8737_{\pm 0.004}$	$0.9002_{\pm 0.005}$	$0.9248_{\pm 0.004}$
LargeKitchenAppliances	$0.8169_{\pm 0.009}$	0.6373	0.696	$0.7342_{\pm 0.007}$	$0.8053_{\pm 0.005}$	0.8729 ±0.003	$0.816_{\pm 0.003}$	$0.8702_{\pm 0.002}$	$0.7556_{\pm 0.009}$	$0.8133_{\pm 0.011}$	$0.7662_{\pm 0.006}$
Lightning2	$0.7322_{\pm 0.009}$	0.6721	0.7049	$0.7869_{\pm 0.033}$	$0.7432_{\pm 0.009}$	$0.7268_{\pm 0.009}$	$0.7596_{\pm 0.009}$	$0.7432_{\pm 0.025}$	$0.6776_{\pm 0.009}$	0.7923 ±0.009	$0.7268_{\pm 0.009}$
Lightning7	0.7671 ±0.014	0.6986	0.7397	$0.7352_{\pm 0.029}$	$0.7352_{\pm 0.016}$	$0.653_{\pm 0.021}$	$0.7489_{\pm 0.029}$	$0.7306_{\pm 0.016}$	$0.6575_{\pm 0.027}$	$0.7397_{\pm 0.027}$	$0.7215_{\pm 0.044}$
Mallat	$0.9555_{\pm 0.009}$	0.9689	0.9484	$0.9164_{\pm 0.016}$	$0.9166_{\pm 0.027}$	$0.9016_{\pm 0.009}$	$0.9032_{\pm 0.007}$	$0.8387_{\pm 0.005}$	$0.8311_{\pm 0.005}$	$0.8708_{\pm 0.01}$	$0.8887_{\pm 0.002}$
Meat	$0.9222_{\pm 0.01}$	0.9833	0.9333	$0.9222_{\pm 0.01}$	$0.8611_{\pm 0.01}$	$0.9333_{\pm 0.017}$	$0.8389_{\pm 0.038}$	$0.8667_{\pm 0.0}$	$0.9278_{\pm 0.019}$	$0.8944_{\pm 0.01}$	$0.9111_{\pm 0.025}$

Appendix

Table 9: Zero-shot feature extraction results on UCR (Random Forest). Second Part.

	Catch22+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
ArticularyWordRecognition	$0.96_{\pm 0.007}$	0.93	0.91	$0.9844_{\pm 0.002}$	$0.99_{\pm 0.0}$	$0.9933_{\pm 0.003}$	$0.9833_{\pm 0.003}$	$0.98_{\pm 0.003}$	$0.9911_{\pm 0.004}$	0.9956 ±0.002	$0.9922_{\pm 0.002}$
BasicMotions	1.0 ±0.0	1.0	0.95	$0.975_{\pm 0.0}$	$0.9667_{\pm 0.014}$	1.0 ±0.0	$0.9833_{\pm 0.014}$	1.0 ±0.0	1.0 ±0.0	1.0 ±0.0	1.0 ±0.0
CharacterTrajectories	0.9814 ±0.001	0.9673	0.9694	$0.9698_{\pm 0.002}$	$0.9603_{\pm 0.0}$	$0.9631_{\pm 0.001}$	$0.9373_{\pm 0.002}$	$0.9517_{\pm 0.002}$	$0.967_{\pm 0.002}$	$0.9761_{\pm 0.001}$	$0.9761_{\pm 0.0}$
Cricket	$0.9444_{\pm 0.014}$	0.8194	0.8472	$0.9722_{\pm 0.0}$	$0.9954_{\pm 0.008}$	$0.9861_{\pm 0.0}$	1.0 ±0.0	$0.9954_{\pm 0.008}$	$0.9907_{\pm 0.008}$	$0.9861_{\pm 0.0}$	$0.9861_{\pm 0.0}$
DuckDuckGeese	$0.5133_{\pm 0.023}$	0.32	0.28	$0.4467_{\pm 0.042}$	$0.42_{\pm 0.02}$	$0.4133_{\pm 0.023}$	$0.4333_{\pm 0.023}$	$0.44_{\pm 0.04}$	$0.4267_{\pm 0.023}$	$0.5067_{\pm 0.031}$	0.54 ±0.02
ERing	$0.8827_{\pm 0.022}$	0.8889	0.8519	$0.9679_{\pm 0.008}$	$0.9765_{\pm 0.008}$	0.9877 ±0.004	$0.9815_{\pm 0.004}$	$0.963_{\pm 0.007}$	$0.9765_{\pm 0.004}$	0.9877 ±0.006	0.9877 ±0.002

Appendix

Table 10: Zero-shot feature extraction results on UEA-27 (Random Forest).

	Catch22+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
ACSF1	0.7	0.8	0.81	0.75	0.75	0.86	0.83	0.83	0.7	0.79	0.76
Adiac	0.665	0.8031	0.8031	0.7775	0.7826	0.8286	0.7187	0.7059	0.8005	0.8414	0.8363
AllGestureWiimoteX	0.5486	0.6229	0.5043	0.6614	0.7014	0.6843	0.67	0.6943	0.6071	0.7171	0.73
AllGestureWiimoteY	0.5229	0.6329	0.5114	0.7029	0.7314	0.7043	0.7371	0.7229	0.6529	0.7414	0.7471
AllGestureWiimoteZ	0.4829	0.5329	0.4529	0.6057	0.6343	0.66	0.6771	0.6586	0.5414	0.6657	0.6814
ArrowHead	0.7143	0.7543	0.7429	0.7771	0.7829	0.8514	0.8229	0.8229	0.7257	0.8457	0.8514

Appendix

Table 11: Zero-shot feature extraction results on UCR (Logistic Regression). First Part.

	Catch22+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
ItalyPowerDemand	0.9388	0.9699	0.9631	0.9543	0.9602	0.964	0.9407	0.9378	0.9116	0.9281	0.9456
LargeKitchenAppliances	0.8373	0.6373	0.696	0.76	0.8293	0.8613	0.872	0.8933	0.7707	0.824	0.816
Lightning2	0.7541	0.6721	0.7049	0.7869	0.7377	0.7869	0.8197	0.8361	0.7377	0.7705	0.7869
Lightning7	0.6986	0.6986	0.7397	0.7397	0.8082	0.7945	0.8493	0.7671	0.7808	0.8356	0.863
Mallat	0.8866	0.9689	0.9484	0.9186	0.9548	0.9467	0.9765	0.9642	0.8486	0.9049	0.9198
Meat	0.9333	0.9833	0.9333	1.0	0.9	0.9333	0.8167	0.85	0.9333	0.9833	0.9667

Appendix

Table 12: Zero-shot feature extraction results on UCR (Logistic Regression). Second Part.

	Catch22+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
ArticularyWordRecognition	0.9333	0.93	0.91	0.9867	0.99	0.9933	0.9833	0.9767	0.9933	0.9933	0.9933
BasicMotions	1.0	1.0	0.95	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0
CharacterTrajectories	0.984	0.9673	0.9694	0.984	0.9694	0.9742	0.9659	0.9714	0.9847	0.9812	0.9868
Cricket	0.9028	0.8194	0.8472	0.9861	1.0	1.0	1.0	1.0	1.0	0.9861	0.9861
DuckDuckGeese	0.42	0.32	0.28	0.48	0.38	0.34	0.48	0.56	0.5	0.54	0.5
ERing	0.9074	0.8889	0.8519	0.9704	0.9778	0.9778	0.9852	0.9815	0.9815	0.9926	0.9889

Appendix

Table 13: Zero-shot feature extraction results on UEA-27 (Logistic Regression).

	Catch22+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
Ego4D	$0.3719_{\pm 0.0}$	NaN	NaN	$0.4324_{\pm 0.0}$	$0.5267_{\pm 0.0}$	$0.5216_{\pm 0.0}$	0.1394	0.1288	$0.5248_{\pm 0.0}$	$0.5519_{\pm 0.0}$	0.5551 ±0.0
EMOPain	$0.7915_{\pm 0.0}$	0.7831	0.7831	$0.8225_{\pm 0.0}$	$0.8254_{\pm 0.0}$	$0.8338_{\pm 0.0}$	0.8507 ±0.0	$0.8028_{\pm 0.0}$	$0.7831_{\pm 0.0}$	$0.8451_{\pm 0.0}$	$0.8169_{\pm 0.0}$
HHAR-ID	$0.9111_{\pm 0.0}$	0.8938	0.9073	$0.9491_{\pm 0.0}$	$0.9708_{\pm 0.0}$	$0.9699_{\pm 0.0}$	$0.9661_{\pm 0.0}$	$0.9693_{\pm 0.0}$	$0.9751_{\pm 0.0}$	$0.9871_{\pm 0.0}$	0.9883 ±0.0
HHAR-OOD	$0.4085_{\pm 0.0}$	0.5311	0.5441	$0.3329_{\pm 0.0}$	$0.3338_{\pm 0.0}$	$0.3691_{\pm 0.0}$	$0.318_{\pm 0.0}$	$0.3222_{\pm 0.0}$	$0.5084_{\pm 0.0}$	$0.4907_{\pm 0.0}$	0.5604 ±0.0
MP8	$0.642_{\pm 0.0}$	0.6235	0.6185	$0.595_{\pm 0.0}$	$0.6134_{\pm 0.0}$	$0.6387_{\pm 0.0}$	$0.6403_{\pm 0.0}$	$0.642_{\pm 0.0}$	$0.6084_{\pm 0.0}$	0.6588 ±0.0	$0.6521_{\pm 0.0}$
MP50	$0.4471_{\pm 0.0}$	0.5664	0.5042	$0.7294_{\pm 0.0}$	0.7563 ±0.0	$0.7345_{\pm 0.0}$	$0.7193_{\pm 0.0}$	$0.6857_{\pm 0.0}$	$0.6992_{\pm 0.0}$	$0.6924_{\pm 0.0}$	$0.716_{\pm 0.0}$

Appendix

Table 14: Zero-shot feature extraction results on HAR (Logistic Regression).

	Catch22+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2
Blink	0.96	0.9178	0.8978	0.9867	0.9889	1.0	0.9889	0.9822	0.5911	0.9489	0.9733
CAP	0.6415	NaN	NaN	0.7374	0.801	0.8168	0.8139	0.815	0.7767	0.8078	0.7945
CAP-OOD	0.6704	NaN	NaN	0.6946	0.7224	0.7535	0.7223	0.729	0.633	0.6415	0.6391
Epilepsy-EEG	0.947	0.9496	0.9447	0.9366	0.9563	0.9414	0.9322	0.9186	0.907	0.9424	0.9312
FingerMovements	0.52	0.5	0.49	0.57	0.62	0.6	0.57	0.53	0.61	0.57	0.64
PCL	0.5402	NaN	NaN	0.5616	0.5667	0.5596	0.5685	0.5731	0.5754	0.575	0.5931

Appendix

Table 15: Zero-shot feature extraction results on EEG (Logistic Regression).

	Catch22+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2	SE-Mantis+	SE-MantisV2	MantisV2 &	MantisV2 &	MantisV2-FT
														TiViT-H	TiConvNext
ACSF1	$0.8233_{\pm 0.021}$	0.8	0.81	0.75	0.75	0.86	0.83	0.83	0.7	0.79	0.76	0.8	0.77	0.81	0.82	$0.77_{\pm 0.01}$
Adiac	$0.734_{\pm 0.007}$	0.8031	0.8031	0.7775	0.7826	0.8286	0.7187	0.7059	0.8005	0.8414	0.8363	0.8312	0.8465	0.7519	0.7749	$0.8321_{\pm 0.007}$
AllGestureWiimoteX	$0.5957_{\pm 0.01}$	0.6229	0.5043	0.6614	0.7014	0.6843	0.67	0.6943	0.6071	0.7171	0.73	0.69	0.6871	0.7229	0.7357	0.779 ±0.006
AllGestureWiimoteY	$0.6319_{\pm 0.016}$	0.6329	0.5114	0.7029	0.7314	0.7043	0.7371	0.7229	0.6529	0.7414	0.7471	0.73	0.7271	0.7557	0.7486	0.7905 ±0.004
AllGestureWiimoteZ	$0.5462_{\pm 0.005}$	0.5329	0.4529	0.6057	0.6343	0.66	0.6771	0.6586	0.5414	0.6657	0.6814	0.6329	0.6357	0.6886	0.7043	0.7505 ±0.017

Appendix

Table 16: The final comparison on UCR. First Part.

	Catch22+	TabPFN	TabICL	MOMENT	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2	SE-Mantis+	SE-MantisV2	MantisV2 &	MantisV2 &	MantisV2-FT
														TiViT-H	TiConvNext
ItalyPowerDemand	$0.9537_{\pm 0.002}$	0.9699	0.9631	0.9543	0.9602	0.964	0.9407	0.9378	0.9116	0.9281	0.9456	0.9436	0.9475	0.9427	0.9446	$0.9517_{\pm 0.001}$
LargeKitchenAppliances	$0.8169_{\pm 0.009}$	0.6373	0.696	0.76	0.8293	0.8613	0.872	0.8933	0.7707	0.824	0.816	0.832	0.816	0.8507	0.888	$0.8596_{\pm 0.008}$
Lightning2	$0.7322_{\pm 0.009}$	0.6721	0.7049	0.7869	0.7377	0.7869	0.8197	0.8361	0.7377	0.7705	0.7869	0.8361	0.7705	0.8525	0.8033	$0.847_{\pm 0.025}$
Lightning7	$0.7671_{\pm 0.014}$	0.6986	0.7397	0.7397	0.8082	0.7945	0.8493	0.7671	0.7808	0.8356	0.863	0.863	0.8219	0.8767	0.7945	$0.7717_{\pm 0.032}$
Mallat	$0.9555_{\pm 0.009}$	0.9689	0.9484	0.9186	0.9548	0.9467	0.9765	0.9642	0.8486	0.9049	0.9198	0.9339	0.9663	0.974	0.9697	$0.9512_{\pm 0.021}$

Appendix

Table 17: The final comparison on UCR. Second Part.

	TS2Vec	T-Loss	TNC	TS-TCC	CNN	Encoder	FCN	MLP	ResNet	TWIESN	DTW	Official	Improved	Catch22+	TabPFN	TabICL	TiRex	Chronos2	TiViT-H	TiConvNext	NuTime	Mantis+	MantisV2	SE-Mantis+	SE-MantisV2	MantisV2 &	MantisV2 &
												MOMENT	MOMENT													TiViT-H	TiConvNext
Adiac	0.762	0.675	0.767	0.767	0.393	0.318	0.841	0.391	0.833	0.428	0.604	0.688	0.7775	0.734	0.8031	0.8031	0.7826	0.8286	0.7187	0.7059	0.8005	0.8414	0.8363	0.8312	0.8465	0.7519	0.7749
AllGestureWiimoteX	0.777	0.763	0.697	0.697	0.411	0.475	0.713	0.477	0.741	0.522	0.716	0.607	0.6614	0.5957	0.6229	0.5043	0.7014	0.6843	0.67	0.6943	0.6071	0.7171	0.73	0.69	0.6871	0.7229	0.7357
AllGestureWiimoteY	0.793	0.726	0.741	0.741	0.479	0.509	0.784	0.571	0.794	0.6	0.729	0.666	0.7029	0.6319	0.6329	0.5114	0.7314	0.7043	0.7371	0.7229	0.6529	0.7414	0.7471	0.73	0.7271	0.7557	0.7486
AllGestureWiimoteZ	0.746	0.723	0.689	0.689	0.375	0.396	0.692	0.439	0.726	0.516	0.643	0.537	0.6057	0.5462	0.5329	0.4529	0.6343	0.66	0.6771	0.6586	0.5414	0.6657	0.6814	0.6329	0.6357	0.6886	0.7043
ArrowHead	0.857	0.766	0.737	0.737	0.717	0.63	0.843	0.784	0.838	0.689	0.703	0.743	0.7771	0.741	0.7543	0.7429	0.7829	0.8514	0.8229	0.8229	0.7257	0.8457	0.8514	0.84	0.8514	0.88	0.8457

Appendix

Table 18: Comparison with more baselines (taken from Goswami et al., 2024 ) on 91 UCR datasets.

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

Feb 21, 2025

Mantis: Lightweight Foundation Model for Time Series Classification

While foundation models have revolutionized various domains, their application to time series classification remains rather under-explored, with existing literature predominantly focused on forecasting. To bridge this gap, we introduce Mantis, a transformer-based foundation model pre-trained exclusively on synthetic data via self-supervised contrastive learning. We demonstrate that effective tokenization is critical to unlocking the full potential of transformers, proposing a novel token generator unit. Furthermore, we introduce an enhanced test-time methodology that bridges the performance gap between Mantis and strong specialized approaches by leveraging intermediate-layer representations, self-ensembling, and cross-model embedding fusion. Extensive experiments demonstrate that Mantis establishes a new state-of-the-art, outperforming existing foundation models across four diverse dataset collections covering various application domains.

Vasilii Feofanov, Songkang Wen, Shifeng Xie +10

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

Time Series Classification Self-Supervised Time Series Representation Learning Time Series Tokenization Time-Series Foundation Models

Ask PDF

Mar 2, 2026

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

Self-supervised foundation models have achieved remarkable success across domains, including time series. However, the potential of non-contrastive methods, a paradigm that has driven significant advances in computer vision, remains underexplored for time series. In this work, we adapt DINOv2-style self-distillation to pretrain a time series foundation model, building on the Mantis tokenizer and transformer encoder architecture as our backbone. Through a student-teacher framework, our method Utica learns representations that capture both temporal invariance via augmented crops and fine-grained local structure via patch masking. Our approach achieves state-of-the-art classification performance on both UCR and UEA benchmarks. These results suggest that non-contrastive methods are a promising and complementary pretraining strategy for time series foundation models.

Yessin Moakher, Youssef Attia El Hili, Vasilii Feofanov

Ecole Polytechnique, France · Huawei Noah’s Ark Lab

Self-Supervised Time Series Representation Learning Time-Series Foundation Models Time Series Classification Self-Supervised Learning Self-Distillation

Ask PDF

Jan 31, 2026

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

The zero-shot evaluation of time series foundation models (TSFMs) for classification typically uses a frozen encoder followed by a task-specific classifier. However, this practice violates the training-free premise of zero-shot deployment and introduces evaluation bias due to classifier-dependent training choices. To address this issue, we propose TIC-FM, an in-context learning framework that treats the labeled training set as context and predicts labels for all test instances in a single forward pass, without parameter updates. TIC-FM pairs a time series encoder and a lightweight projection adapter with a split-masked latent memory Transformer. We further provide theoretical justification that in-context inference can subsume trained classifiers and can emulate gradient-based classifier training within a single forward pass. Experiments on 128 UCR datasets show strong accuracy, with consistent gains in the extreme low-label situation, highlighting training-free transfer for time series classification.The source code is publicly available at https://github.com/fangjuntao/TIC-FM.

Juntao Fang, Shifeng Xie, Shengbin Nie +7

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

Time Series Classification In-Context Learning Time-Series Foundation Models

Ask PDF

Abstract

Time Series Classification

Explore similar work

Mantis: Lightweight Foundation Model for Time Series Classification

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

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