Abstract: Recurrent architectures have long been the natural choice for modeling time-dependent data, yet the recent dominance of Transformers has shifted attention away from their unique advantages. In this talk, I will present TiRex, a new zero-shot time series foundation model built on xLSTM, an enhanced recurrent architecture that combines the state-tracking power of LSTMs with in-context learning capabilities previously seen only in Transformers. TiRex leverages these properties to deliver state-of-the-art zero-shot forecasting results on the HuggingFace GiftEval and Chronos-ZS benchmarks, where it outperforms significantly larger Transformer-based models from Google, Amazon, Salesforce, and Alibaba. Beyond its benchmark results, TiRex demonstrates that recurrent models cannot only rival but Pareto-dominate Transformers in accuracy, efficiency, and interpretability, suggesting a new generation of sustainable, foundation models for temporal intelligence.

Register Now

Sepp Hochreiter is Head of the Institute for Machine Learning, the LIT AI Lab, and Director of the ELLIS Unit at Johannes Kepler University Linz. He is also the founder of NXAI, a company driving innovation in deep learning with technologies like xLSTM and the time series foundation model TiRex.

Widely recognized as a pioneer of deep learning, Hochreiter is best known for his foundational contributions, including the development of Long Short-Term Memory (LSTM) networks and the identification of the vanishing gradient problem. LSTMs laid the groundwork for the first large language models and were instrumental in the evolution of systems like ChatGPT.