(IJCAI'26) From Time Series Analysis to Question Answering: A Survey in the LLM Era
✨ If you find our work useful for your research, please consider giving it a star ⭐ on GitHub to stay updated with future releases.
🌐 Application
We applied our taxonomy to a literature interpretation tool and provided an Example that retrieves the latest relevant literature, helping readers quickly understand related works. We also provide a general ScholarAgent for subscription and use.
📚 Paper
Taxonomy of relevant literature across three alignment paradigms. The icon to the left of each method indicates its domain, while the absence of this icon denotes a general domain. The icon on the right denotes the use of time series visual plots. The background color of each method represents task targets: lighter shades denote time series analysis (TSA), and darker shades denote time series question answering (TSQA). The TSA literature dominates external Injective and Bridging Alignment, whereas the TSQA literature dominates Internal Alignment, reflecting the evolution from TSA to TSQA driven by a shift from external to internal alignment. Compared with existing surveys, this evolution perspective integrates previously isolated TSA and TSQA research into a unified taxonomy of alignment paradigms.
Alignment Paradigm
The horizontal axis indicates whether LLM parameters are trained, and the vertical axis indicates whether temporal modifications are required. Temporal modification refers to adapting time series for LLMs, including both modifications outside the LLM and adjustments to the internal LLM architecture. These two dimensions determine the three alignment paradigms.
(a) Injective Alignment involves no temporal modification and adopts frozen LLMs. This design preserves the original LLM parameters.
(b) Bridging Alignment introduces temporal modification while still employing frozen LLMs. This design enables joint processing of time series and textual inputs while preserving all parameters of the original LLM.
(c) Internal Alignment combines temporal modification with training LLMs by parameter updating to provide native support for time series.
We organize existing literature into the three alignment paradigms, along with related surveys and datasets.
Taxonomy
Injective
Bridging:
Internal:
Survey:
Dataset:
| Date | Paper | Institute | Publication | Domain | LLMs |
|---|---|---|---|---|---|
| 7 Nov 2025 |
QuAnTS: Question Answering on Time Series[Code] | TU Darmstadt | Preprint | IoT | Llama3.1-8B |
| 28 Jun 2025 |
Time-MQA: Time Series Multi-Task Question Answering with Context Enhancement[Code] | University of Oxford | ACL'25 | General | Mistral-7B |
| 25 Jun 2025 |
ITFormer: Bridging Time Series and Natural Language for Multi-Modal QA with Large-Scale Multitask Dataset[Code] | Shanghai Jiao Tong University | ICML'25 | General | GPT-4o, Gemini-Pro |
| 5 Jun 2025 |
Context is Key: A Benchmark for Forecasting with Essential Textual Information (CiK)[Code] | ServiceNow Research | ICML'25 | General | Qwen-2.5-7B, Llama-3-70B, Llama-3.1-405B |
| 21 Mar 2025 |
MTBench: A Multimodal Time Series Benchmark for Temporal Reasoning and Question Answering[Code] | Yale University | Preprint | Financial | GPT-4o, Gemini, Claude, DeepSeek, Llama3.1 |
| 13 Mar 2025 |
Chat-TS: Enhancing Multi-Modal Reasoning Over Time-Series and Natural Language Data | Queen’s University | Preprint | General | LLama 3.1-8B |
| 3 Mar 2025 |
SensorQA: A Question Answering Benchmark for Daily-Life Monitoring[Code] | Georgia Institute of Technology | SenSys'25 | IoT | GPT-3.5-Turbo, GPT-4-Turbo |
| 18 Oct 2024 |
TimeSeriesExam: A time series understanding exam[Code] | Carnegie Mellon University | NeurIPS'24 Workshop | General | GPT-4o, Gemini, Phi3.5 |
| 12 Jun 2024 |
Time-MMD: Multi-Domain Multimodal Dataset for Time Series Analysis[Code] | Georgia Institute of Technology | NeurIPS'24 | General | LLaMA-3, GPT-2 |
| 17 Apr 2024 |
Language Models Still Struggle to Zero-shot Reason about Time Series (TSandLanguage)[Code] | University of Washington | EMNLP'24 (Findings) | General | GPT-4 |
| 28 Oct 2023 |
Insight Miner: A Time Series Analysis Dataset for Cross-Domain Alignment with Natural Language | University of California, Berkeley | NeurIPS'23 Workshop | General | LLaVA, GPT-4 |
| 27 Oct 2023 |
JoLT: Jointly Learned Representations of Language and Time-Series | Carnegie Mellon University | NeurIPS'23 Workshop | Medical | GPT-2, OPT |
Citation
If you find this useful, please cite our paper: "From Time Series Analysis to Question Answering: A Survey in the LLM Era".
@article{li2026tsa2tsqa,
title={From Time Series Analysis to Question Answering: A Survey in the LLM Era},
author={Wei Li and Zhe Xie and Yuxuan Liang and Xinli Hao and Yunyao Cheng and Dan Pei and Xiaofeng Meng},
journal={arXiv preprint arXiv:2506.11512},
year={2026}
}
Contact Us
For inquiries or further assistance, contact us at [email protected].