**UNITS: A Unified Multi-Task Time Series Model** **Authors:** Shanghua Gao **Abstract:** The shift from conventional deep learning methods to pre-trained foundational models is driving advancements in time series models. While pre-trained transformers and reprogrammed text-based LLMs have achieved state-of-the-art results, the best-performing architectures vary significantly across tasks, and models often have limited scope. Models that unify predictive and generative time series tasks under a single framework remain challenging to achieve. This paper introduces UNITS, a multi-task time series model that uses task tokenization to express predictive and generative tasks within a single model. UNITS leverages a modified transformer block designed to obtain universal time series representations, enabling transferability from heterogeneous, multi-domain pre-training datasets to various downstream datasets. Across 38 datasets spanning human activity sensors, healthcare, engineering, and finance domains, UNITS outperforms 12 forecasting models, 20 classification models, 18 anomaly detection models, and 16 imputation models, including repurposed text-based LLMs. UNITS demonstrates effective few-shot and prompt learning capabilities when evaluated on new data domains and tasks. In the conventional single-task setting, UNITS outperforms strong task-specialized time series models. **Introduction:** Foundation models, especially large language models (LLMs), have transformed deep learning by supporting multiple tasks with a single pretrained model. This approach is efficient for adapting to new tasks with little to no additional training via multi-task learning, few-shot learning, and prompting. However, achieving a unified multi-task time series model that can handle diverse tasks remains challenging due to the heterogeneity of time series datasets and the fundamentally different objectives of generative and predictive tasks. UNITS addresses these challenges by using task tokenization to convert task specifications into a unified token representation, enabling universal task specification without post hoc modifications to the network architecture. The unified time series architecture processes heterogeneous time series with diverse variables and lengths without modifying the network architecture, using self-attention across both time and variable dimensions. UNITS supports both generative and predictive tasks through a shared weight model, enabling co-training on a wide variety of datasets. **Related Work:** Traditional time series modeling has been extensively explored, but task-specific models are typically trained separately for each dataset, requiring specialized modules. General time series modeling aims to develop models with broad capabilities, similar to foundation models for language and vision. Prompt learning has emerged as an efficient method for task adaptation in large models, but existing approaches often require computationally expensive pre-trained LLMs. UNITS, in contrast, is trained exclusively on time series data and uses universal task tokenization to adapt to new tasks beyond forecasting. **Problem Formulation:** The desiderata for a unified multi-task time series model include handling heterogeneous time series, adopting a universal task specification, and sharing weights across tasks. UNITS supports multi-task, prompt-based, zero**UNITS: A Unified Multi-Task Time Series Model** **Authors:** Shanghua Gao **Abstract:** The shift from conventional deep learning methods to pre-trained foundational models is driving advancements in time series models. While pre-trained transformers and reprogrammed text-based LLMs have achieved state-of-the-art results, the best-performing architectures vary significantly across tasks, and models often have limited scope. Models that unify predictive and generative time series tasks under a single framework remain challenging to achieve. This paper introduces UNITS, a multi-task time series model that uses task tokenization to express predictive and generative tasks within a single model. UNITS leverages a modified transformer block designed to obtain universal time series representations, enabling transferability from heterogeneous, multi-domain pre-training datasets to various downstream datasets. Across 38 datasets spanning human activity sensors, healthcare, engineering, and finance domains, UNITS outperforms 12 forecasting models, 20 classification models, 18 anomaly detection models, and 16 imputation models, including repurposed text-based LLMs. UNITS demonstrates effective few-shot and prompt learning capabilities when evaluated on new data domains and tasks. In the conventional single-task setting, UNITS outperforms strong task-specialized time series models. **Introduction:** Foundation models, especially large language models (LLMs), have transformed deep learning by supporting multiple tasks with a single pretrained model. This approach is efficient for adapting to new tasks with little to no additional training via multi-task learning, few-shot learning, and prompting. However, achieving a unified multi-task time series model that can handle diverse tasks remains challenging due to the heterogeneity of time series datasets and the fundamentally different objectives of generative and predictive tasks. UNITS addresses these challenges by using task tokenization to convert task specifications into a unified token representation, enabling universal task specification without post hoc modifications to the network architecture. The unified time series architecture processes heterogeneous time series with diverse variables and lengths without modifying the network architecture, using self-attention across both time and variable dimensions. UNITS supports both generative and predictive tasks through a shared weight model, enabling co-training on a wide variety of datasets. **Related Work:** Traditional time series modeling has been extensively explored, but task-specific models are typically trained separately for each dataset, requiring specialized modules. General time series modeling aims to develop models with broad capabilities, similar to foundation models for language and vision. Prompt learning has emerged as an efficient method for task adaptation in large models, but existing approaches often require computationally expensive pre-trained LLMs. UNITS, in contrast, is trained exclusively on time series data and uses universal task tokenization to adapt to new tasks beyond forecasting. **Problem Formulation:** The desiderata for a unified multi-task time series model include handling heterogeneous time series, adopting a universal task specification, and sharing weights across tasks. UNITS supports multi-task, prompt-based, zero