✨ If you find our survey useful, a star ⭐ on GitHub helps others discover it and keeps you updated on future releases.
Time series reasoning treats time as a first-class axis and integrates intermediate evidence into the answer itself.
This survey organizes the field along two levels:
-
Reasoning Topology — execution structures:
- Direct reasoning (single step)
- Linear chain reasoning (sequential intermediate steps)
- Branch-structured reasoning (exploration, feedback, and aggregation)
-
Primary Objective — the main intent:
- Traditional time series analysis (forecasting, classification, anomaly detection, segmentation)
- Explanation and understanding (temporal QA, diagnostics, structure discovery)
- Causal inference and decision making (counterfactuals, policy evaluation, decision support)
- Time series generation (simulation, editing, synthesis)
We complement this two-level taxonomy with attribute tags that capture control-flow operators (decomposition, verification, ensembling), execution actors (tool use, agents), information sources (knowledge access, multimodality), and alignment regimes (prompting, supervised finetuning, reinforcement/preference, hybrid).
This repository curates and classifies both research papers and non-research contributions in the field.
- Research papers are categorized by reasoning topology and primary objective.
- Non-research papers include datasets & benchmarks (reasoning-first, reasoning-ready, general-purpose TS), as well as surveys, tutorials, and position/vision papers.
The following table presents the unified taxonomy and tags for all research papers we review, while non-research contributions are grouped separately for completeness.
The distribution of surveyed papers, split into research and non-research categories:
- Abbreviations & Value Definitions
- Direct Reasoning
- Linear Chain Reasoning
- Branch-Structured Reasoning
- Non-Research Papers
Covers: abbreviations and value definitions for primary objectives, task values, and attribute tags used in the curated research-paper taxonomy tables.
Non-research papers are listed without attribute tags.
| Full Name | Abbreviation |
|---|---|
| Task Decomposition | T-Dec |
| Verification and Critique | T-Ver |
| Ensemble Selection | T-Ens |
| Tool Use | T-Tool |
| Knowledge Access | T-Know |
| Multimodal Inputs | T-Multi |
| Agents | T-Agent |
| LLM Alignment | T-Align |
| Full Name | Abbreviation |
|---|---|
| Traditional Time Series Analysis | Trad. TS Anal. |
| Explanation and Understanding | Expl. & Und. |
| Causal Inference and Decision Making | Causal Inf. |
| Time Series Generation | TS Gen. |
| Full Name | Abbreviation |
|---|---|
| Forecasting | Forc. |
| Classification | Class. |
| Anomaly Detection | Anom. Det. |
| Segmentation | Segm. |
| Multiple Tasks | Mult. Tasks |
| Temporal Question Answering | Temp. QA |
| Explanatory Diagnostics | Expl. Diagn. |
| Structure Discovery | Struct. Disc. |
| Autonomous Policy Learning | Auto. Policy |
| Advisory Decision Support | Adv. Dec. Supp. |
| Conditioned Synthesis | Cond. Synth. |
| Tag(s) | Values / Meaning |
|---|---|
| T-Dec, T-Ver, T-Ens, T-Tool, T-Know, T-Multi | ✔ = present, empty = absent |
| T-Agent | 0 = no agent, 1 = single agent, M = multiple agents |
| T-Align | P = Prompting, S = Supervised fine-tuning, R = Reinforcement/preference alignment, H = Hybrid |
We welcome contributions to keep this table updated. Please follow these steps:
-
Decide Research vs. Non-Research
- Research = new methods/analyses with experiments (papers with tasks, tags, etc.).
- Non-Research = datasets/benchmarks, surveys/tutorials, position/vision papers.
-
If Research:
- Pick exactly one Reasoning Topology and add the paper to that table:
- Direct Reasoning → Research Papers for Direct Reasoning
- Linear Chain Reasoning → Research Papers for Linear Chain Reasoning
- Branch-Structured Reasoning → Research Papers for Branch-Structured Reasoning
- Fill columns: Primary Objective, Task, and attribute tags
(T-Dec,T-Ver,T-Ens,T-Tool,T-Know,T-Multi,T-Agent∈ {0,1,M},T-Align∈ {P,S,R,H}). - One objective + one topology only. Keep tags consistent with the legends.
- Pick exactly one Reasoning Topology and add the paper to that table:
-
If Non-Research:
- Add the paper under Non-Research Papers in the correct subgroup:
- Reasoning-First Benchmarks
- Reasoning-Ready Benchmarks
- General-Purpose Time Series Benchmarks
- Surveys and Tutorials
- Position and Vision Papers
- No attribute tags for non-research entries.
- Add the paper under Non-Research Papers in the correct subgroup:
-
Formatting rules
- Use the exact row format already in the tables:
| [Paper Title](link) [Venue/Year] | <Type or columns per table> | - Keep venue/year brackets (e.g.,
[NeurIPS 2024],[arXiv 2025]). - Place new rows in chronological order (newest first) within each subgroup/table.
- Use the exact row format already in the tables:
-
Open a PR
- Title:
Add <Paper Short Title> (<Year>) - In the PR body, state:
- Research vs. non-research
- Chosen reasoning topology (if research)
- Primary objective, task, and tags you set
- Any notes (e.g., multimodal inputs, tools used)
- Title:
If you find this resource useful, please cite our survey.
@misc{chang2025surveyreasoningagenticsystems,
title={A Survey of Reasoning and Agentic Systems in Time Series with Large Language Models},
author={Ching Chang and Yidan Shi and Defu Cao and Wei Yang and Jeehyun Hwang and Haixin Wang and Jiacheng Pang and Wei Wang and Yan Liu and Wen-Chih Peng and Tien-Fu Chen},
year={2025},
eprint={2509.11575},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2509.11575},
}