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

In plain words

Sensors, bank ledgers, traffic counters and hospital records all produce data that changes over time, and often over space too: a road network where each junction's speed depends on its neighbours. For years these problems were handled by small models built for one task each. Large language models and other very large pre-trained models have since shown that one general model can often be reused across many tasks, so researchers have started to apply them to this kind of data.

This survey is a map of that work. It asks whether large models can handle time series and data that has both space and time, and it sorts the existing papers along a few questions. Is the data a plain time series, or a graph, a knowledge graph of dated facts, or a video? Is the model a language model that has been adapted, or a model built for temporal data from the start? Is it meant to work across many domains, or for one field such as traffic, finance or healthcare? The authors then list the datasets and software tools used in each area, and close with open problems: theory, multimodal inputs, continual learning, explanation, privacy and safe decision-making.

Read it as a reading list and a vocabulary for the field. It does not propose a new model and it does not benchmark anything.

TL;DR

A survey with no experiments and no new model. It organises the field into two branches, time series (LM4TS) and spatio-temporal data (LM4STD) §3 fig. 3. Each branch is split into large language models (LLMs) and pre-trained foundation models (PFMs), and then into general-purpose and domain-specific work §3. The spatio-temporal branch covers graphs, temporal knowledge graphs and video §5. Methods are listed with their category, scope, task, venue and year in tab. 2; datasets, models and tools by application are in tab. 3. The authors argue that earlier surveys rarely cover these data types together, and their comparison table marks their own paper in every column tab. 1.

Problem

Large models are reshaping many fields, but time series and spatio-temporal data have been modelled mostly with statistical methods and small task-specific networks. The authors argue that these data types deserve a joint review, because both are temporal and share modelling methods, and because the gaps between them can transfer ideas §1. The introduction names data scarcity in many domains as a limit on large-model training §1.

The survey's stated scope covers time series, spatio-temporal graphs, temporal knowledge graphs and video §1 §2.3. It excludes works where large models are only incidental, conventional small networks without pre-training or adaptation, and duplicate versions without new technical content §1, "Survey Methodology and Scope".

Method

The survey proposes a taxonomy. The steps below reproduce how the authors place a paper in it, based on the definitions in §2 and §3.

1. Data category. Time series (LM4TS) or spatio-temporal data (LM4STD) §3. Time series are ordered sequences that may be univariate or multivariate §2.3 def. 1. Spatio-temporal data is organised in space and time §2.3; the survey keeps three forms: spatio-temporal graphs (STGs) §2.3 def. 2, temporal knowledge graphs (TKGs) §2.3 def. 3, and video §2.3 def. 4 §5.
2. Model type. LLM or PFM §2 §3.
   • LLM: a language-centred foundation model repurposed for temporal or spatio-temporal tasks through prompting, reprogramming or adaptation. The authors count a method as an LLM approach when the language model is a central backbone, interface or reasoning module, whether or not it was trained further §2.
   • PFM: a pre-trained model designed or adapted as a general backbone for temporal, spatial, relational, visual or multimodal data §2.2 §3.
3. Scope. General-purpose or domain-specific §3. Domains named for time series are transportation, finance and healthcare §4.1.2 §4.2.2. For STD, the authors sort by domain or modality §3.
4. Task. Time series tasks are forecasting, classification, anomaly detection and imputation §2.3. STG tasks are mainly forecasting, plus link prediction and node or graph classification §2.3. TKG tasks are completion and forecasting §2.3. Video tasks are detection, captioning, anticipation and querying §2.3.

# Taxonomy placement of one surveyed paper  [§3] [fig. 3]
# Input:  a paper's data type, backbone and application domain
# Output: a (family, model type, scope, task) label

def categorise(paper):
    # 1. Split by data category: time series or spatio-temporal  [§3]
    if paper.data_type == "time series":
        family = "LM4TS"
    else:
        family = "LM4STD"
    # 2. Split by model type: language-centred or general backbone  [§2]
    if paper.backbone_is_language_model:
        model_type = "LLM"
    else:
        model_type = "PFM"
    # 3. Split by scope: many domains or one domain  [§3]
    if paper.target_domains == "many":
        scope = "general-purpose"
    else:
        scope = "domain-specific"
    # 4. Record the task from the task list for the data type  [§2.3]
    task = paper.primary_task
    return family, model_type, scope, task

The survey's own text describes two cross-cutting challenges that recur in each branch. The first is turning temporal observations into tokens a model can read, such as patches, symbols, summaries or prompt-aligned embeddings for time series, and extra structure such as topology, entity relations or video frames for STD §5.4. The second is handling long dependencies within a limited context, which the authors address with patching, decomposition or adapters for time series, and with graph summarisation, pooling, memory and frame selection for STD §5.4.

Hyperparameters

Not applicable. The survey trains and tunes no model, so it states no hyperparameters §1. The settings of the surveyed methods are in their own papers; this card does not restate them.

Results

No experiments. The survey reports no benchmark numbers of its own. Its outputs are the coverage tables:

Coverage numbers, such as how many papers were reviewed, are not stated in the text.

Author-reported: the survey's claim of a unified and up-to-date taxonomy and its breadth. Compare with the cited surveys in tab. 1 before relying on the claim.

Paper vs code

No implementation code. The survey describes no model of its own.

The arXiv listing links one repository, the authors' curated list of papers and resources on large models for time series and spatio-temporal data README.md. The clone holds only a README (not read in full) and no code or LICENSE file, so the licence is not stated. The card's code field points to this list only because the authors' listing links it; it is not an implementation. Search performed: the arXiv abstract page (comments field), the repository linked there, and one GitHub repository search. No other official code was identified.

Table 3 links third-party dataset and tool repositories, for example a PEMS04 link to a repository of a different project tab. 3. These were not checked here, and dataset licences are not stated in the paper.

Pitfalls

• Scope and inclusion are described only in prose. No search dates, numbers of records screened or inclusion counts are given §1, "Survey Methodology and Scope". Use the table entries as examples, not as a complete census.
• The LLM/PFM boundary is soft. The authors call the split practical rather than ontological, and note that PFM4TS is nascent §3. A method can fall into either group depending on how the language model is used; follow the placement in tab. 2, not a rule.
• Broader data types are defined but not covered. The background defines point cloud streams and trajectories, but the survey limits itself to STGs, TKGs and video §2.3.
• Video claims lean on general-purpose models. The video introduction presents CLIP and DALL-E as models that extract contextual information from video §5.3. Neither is a video-specific model in that paragraph. The card does not check these claims against the cited sources.
• Some outlook points rest on single examples. The discussion of video anomaly detection networks §7.2 and of battery state-of-health estimation §7.3 generalises from one cited work each.
• Self-assessment. Table 1 gives the survey itself full marks across all its columns, with no criteria stated for marking tab. 1.
• Licences of listed resources. Table 3 lists datasets and tools without their licences. Before reusing any of them, check each source; this card does not give licence terms for them.
• Time-sensitive content. The survey covers work up to its submission date, which the text does not state. This card is built from the LaTeX source fetched on 2026-10-18; the arXiv record lists version v3, updated 2026-06-08. Work published after the survey is absent.

Lineage

• Covers and compares: the prior surveys in tab. 1, among them Ma et al., Latif et al., Madan et al. and Liang et al. tab. 1.
• Traces two lines of prior work: the time series foundation-model line, exemplified by Chronos and Time-MoE §1 §4.2.1, and the prompting and reprogramming line, exemplified by PromptCast and Time-LLM §4.1.1.
• Superseded by: not known from this card.