Mantis: Lightweight Foundation Model for Time Series Classification

Vasilii Feofanov, Songkang Wen, Shifeng Xie, Simon Roschmann, Marius Alonso, Hongbo Guo, Romain Ilbert, Malik Tiomoko, Quentin Bouniot, Zeynep Akata, Lujia Pan, Jianfeng Zhang, Ievgen Redko
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:30692-30719, 2026.

Abstract

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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-feofanov26a, title = {Mantis: Lightweight Foundation Model for Time Series Classification}, author = {Feofanov, Vasilii and Wen, Songkang and Xie, Shifeng and Roschmann, Simon and Alonso, Marius and Guo, Hongbo and Ilbert, Romain and Tiomoko, Malik and Bouniot, Quentin and Akata, Zeynep and Pan, Lujia and Zhang, Jianfeng and Redko, Ievgen}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {30692--30719}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/feofanov26a/feofanov26a.pdf}, url = {https://proceedings.mlr.press/v306/feofanov26a.html}, abstract = {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.} }
Endnote
%0 Conference Paper %T Mantis: Lightweight Foundation Model for Time Series Classification %A Vasilii Feofanov %A Songkang Wen %A Shifeng Xie %A Simon Roschmann %A Marius Alonso %A Hongbo Guo %A Romain Ilbert %A Malik Tiomoko %A Quentin Bouniot %A Zeynep Akata %A Lujia Pan %A Jianfeng Zhang %A Ievgen Redko %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-feofanov26a %I PMLR %P 30692--30719 %U https://proceedings.mlr.press/v306/feofanov26a.html %V 306 %X 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.
APA
Feofanov, V., Wen, S., Xie, S., Roschmann, S., Alonso, M., Guo, H., Ilbert, R., Tiomoko, M., Bouniot, Q., Akata, Z., Pan, L., Zhang, J. & Redko, I.. (2026). Mantis: Lightweight Foundation Model for Time Series Classification. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:30692-30719 Available from https://proceedings.mlr.press/v306/feofanov26a.html.

Related Material