Technical Summary: Large Models for Time Series and Spatio-Temporal Data: A Survey and Outlook

1. Problem Statement

Temporal data, encompassing time series and spatio-temporal data, is pervasive in real-world applications ranging from geoscience and transportation to healthcare and finance. While these data types are critical for understanding dynamic system behaviors, their analysis has historically relied on statistical modeling and compact deep learning architectures (e.g., RNNs, CNNs, GNNs, and standard Transformers). These traditional models are often task-specific, limited in their ability to support broad transfer learning, and lack robust semantic representation and multi-task reasoning capabilities.

Although Large Language Models (LLMs) and Pre-trained Foundation Models (PFMs) have achieved remarkable success in Natural Language Processing (NLP), Computer Vision (CV), and other modalities, their application to temporal data remains fragmented. A significant challenge lies in the limited availability of high-quality, diverse large-scale datasets for temporal data compared to text or images. Furthermore, there is a lack of a unified framework that integrates the study of time series and spatio-temporal data (including graphs, knowledge graphs, and videos) under the paradigm of large models. The field requires a comprehensive review to consolidate recent advances, clarify methodological distinctions, and identify open research opportunities.

2. Methodology and Taxonomy

This paper does not propose a new algorithm or experimental model but rather provides a comprehensive survey and systematic taxonomy of existing large models tailored for time series and spatio-temporal data. The authors organize the literature along four primary dimensions: data types, model categories, model scopes, and application areas/tasks.

2.1 Core Definitions

The survey distinguishes between two operational groups of large models:

• Large Language Models (LLMs): Language-centered foundation models repurposed for temporal tasks. This includes both embedding-visible models (open-source, e.g., Llama, BLOOM) and embedding-invisible models (closed-source, e.g., GPT-4, PaLM). Adaptation strategies include prompting, reprogramming, and fine-tuning (e.g., adapters, LoRA).
• Pre-trained Foundation Models (PFMs): Models specifically designed or adapted as general-purpose backbones for temporal, spatial, relational, or multimodal data. These are distinct from task-specific models and aim for emergent capabilities like modality bridging and reasoning.

2.2 Unified Taxonomy

The authors categorize the field into two main clusters:

1. Large Models for Time Series (LM4TS):
   • LLM4TS: Repurposing language models for time series tasks (forecasting, classification, etc.).
   • PFM4TS: Models pre-trained specifically on time series data (e.g., Chronos, TimesFM).
   • Scope: Divided into General-Purpose (cross-domain) and Domain-Specific (Transportation, Finance, Healthcare).
2. Large Models for Spatio-Temporal Data (LM4STD):
   • LLM4STD & PFM4STD: Models handling Spatio-Temporal Graphs (STGs), Temporal Knowledge Graphs (TKGs), and Video Data.
   • Scope: Similarly divided into General-Purpose and Domain-Specific (e.g., Climate, Urban Computing).

2.3 Review Scope

The survey covers works published or released up to the time of submission (circa 2025-2026), sourced from major databases (ACM, IEEE, arXiv, etc.). It includes:

• Time Series: Forecasting, classification, anomaly detection, imputation.
• Spatio-Temporal Graphs: Traffic forecasting, link prediction.
• Temporal Knowledge Graphs: Completion and forecasting.
• Video Data: Question answering, captioning, anticipation.

3. Key Contributions

The paper outlines four primary contributions:

1. A Comprehensive and Foundational Survey: It offers an in-depth review of advancements in large models for temporal data, serving as a foundational resource for researchers and practitioners by systematically charting the field's landscape.
2. Unified and Structured Taxonomy: The authors introduce a novel classification system that organizes research into LM4TS and LM4STD, further subdivided by model type (LLM vs. PFM) and scope (General vs. Domain-Specific). This provides a coherent roadmap for understanding the field from multiple perspectives.
3. Abundant Resource Compilation: The paper curates a wealth of resources, including datasets (e.g., METR-LA, MIMIC-III, CMIP6), open-source implementations, and evaluation benchmarks, organized by major application areas.
4. Future Research Opportunities: The authors identify and elaborate on six promising avenues for future research:
   • Theoretical Analysis: Understanding the transferability of linguistic representations to temporal sequences.
   • Multimodal Models: Aligning multi-resolution temporal signals across text, numbers, and images.
   • Continuous Learning: Addressing non-stationarity and concept drift without catastrophic forgetting.
   • Interpretability: Developing frameworks to explain model decisions in high-stakes domains.
   • Privacy and Security: Mitigating risks of data leakage and adversarial attacks in sensitive temporal data.
   • LLM Agents: Enabling robust, uncertainty-aware decision-making in time series systems.

4. Key Findings and Results

The survey synthesizes the current state of the art without presenting new experimental results, highlighting the following trends and observations:

• Repurposing vs. Pre-training: There is a dual trend where researchers either repurpose existing LLMs (via prompting or reprogramming, e.g., Time-LLM, PromptCast) or develop dedicated PFMs pre-trained on massive temporal corpora (e.g., Chronos, Pangu-Weather).
• Domain Adaptation: Significant progress has been made in domain-specific applications. In Healthcare, models like NYUTron and MedTsLLM show promise in clinical event prediction. In Finance, models like M2VN and GPT4FTS integrate textual news with numerical data for volatility forecasting. In Climate, models like Pangu-Weather and ClimaX have revolutionized global forecasting with high accuracy and reduced computational costs.
• Spatio-Temporal Complexity: For spatio-temporal graphs and videos, large models are increasingly used to integrate textual semantics with structural or visual data. Models like STG-LLM and Video-LLaMA demonstrate the ability to handle complex interactions and long-range dependencies.
• Challenges: The survey notes that despite progress, challenges remain, including the scarcity of high-quality large-scale temporal datasets, the difficulty of aligning symbolic graph dynamics with continuous temporal changes, and the computational costs associated with processing long video sequences or large dynamic networks.

5. Significance and Outlook

The paper positions itself as a catalyst for the burgeoning field of large models for temporal data. Its significance lies in:

• Integration: It bridges the gap between time series and spatio-temporal analysis, fostering cross-pollination of ideas between domains that were previously studied in isolation.
• Standardization: By providing a unified taxonomy, it helps standardize the terminology and classification of models, aiding in the comparison of different approaches.
• Resource Accessibility: The compilation of datasets and tools lowers the barrier to entry for new researchers.
• Direction Setting: By outlining specific future directions (e.g., theoretical analysis of transferability, privacy-preserving PFMs), the paper guides the community toward addressing the fundamental limitations of current methodologies.

The authors conclude that while the integration of large models into temporal data analysis faces unique challenges, the potential for advancing artificial general intelligence (AGI) in understanding and processing temporal dynamics is substantial. The survey aims to stimulate further curiosity and research in this dynamic area.