From Universal Prediction to Universal Workflows: A Survey of Foundation Models for Tabular and Time Series Data
Hao-Run Cai , Jun-Peng Jiang , Si-Yang Liu , Han-Jia Ye
Structured data plays a central role in many critical applications, ranging from finance to healthcare. While Foundation Models have achieved remarkable success in natural language processing and computer vision, extending them to structured data—primarily tabular and time series—remains challenging. Although tabular and time series data exhibit different structural properties, they share several fundamental difficulties, including heterogeneous data schemas, the lack of a fixed vocabulary for continuous values, and substantial distribution shifts across datasets. In response to these challenges, both communities have developed closely related ideas and solutions, despite using different terms and modeling approaches. This survey brings together recent advances in Foundation Models for time series and tabular data under a unified perspective. We systematically organize existing methods along four key dimensions: Input Encoding, Model Architecture, Pre-training Strategies, and Adaptation Techniques, covering both native structured-data foundation models and LLM-adapted approaches. Beyond task-level prediction, we further extend the discussion to process-level generalization by reviewing emerging agentic frameworks that support end-to-end data workflows. Finally, we outline several promising research directions, including omni-modal intelligence and composable ecosystems.
Structured Data / Tabular Data / Time Series / Foundation Model / LLM Agents
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(3396KB)
