Leto: Modeling Multivariate Time Series with Memorizing at Test Time

ICLR 2026 Conference Submission21811 Authors

Keywords: Multivariate Time series, Time Series Forecasting, Time Series Classification, Transformers, Recurrent neural networks

Abstract: Modeling multivariate time series data has been at the forefront of machine learning research efforts across diverse domains. However, effectively capturing dependencies across both time and variate dimensions, as well as temporal dynamics, have made this problem extremely challenging under realistic settings. The recent success of sequence models, such as Transformers, Convolutions, and Recurrent Neural Networks, in language modeling and computer vision tasks, has motivated various studies to adopt them for time series data. These models, however, are either: (1) natively designed for a univariate setup thus missing the the rich information that comes from the inter-dependencies of time and variate dimensions; (2) inefficient for long-range time series; and/or (3) propagating the prediction error over time. In this work, we present Leto, a native 2-dimensional memory module that takes the advantage of temporal inductive bias across time while maintaining the permutation equivariance of variates. Leto uses meta in-context memory modules to learn and memorize patterns across the time dimension, and simultaneously, incorporates information from other correlated variates, if needed. Our experimental evaluation shows the effectiveness of Leto on extensive and diverse benchmarks, including time series forecasting (short, long, and ultra-long), classification, and anomaly detection.

Primary Area: learning on time series and dynamical systems

Submission Number: 21811