🤖 AI

From Time Series Analysis to Question Answering: A Survey in the LLM Era

This survey explores the evolution from traditional Time Series Analysis to Time Series Question Answering in the LLM era, proposing a taxonomy and three alignment paradigms to bridge the gap between language models and temporal data while outlining future research directions.

Original authors: Wei Li, Zhe Xie, Yuxuan Liang, Xinli Hao, Yunyao Cheng, Dan Pei, Xiaofeng Meng

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wei Li, Zhe Xie, Yuxuan Liang, Xinli Hao, Yunyao Cheng, Dan Pei, Xiaofeng Meng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: From "Doing Math" to "Having a Chat"

Imagine you have a massive, complex dashboard of numbers representing the heartbeat of a city, a stock market, or a patient's health. For years, experts have used Time Series Analysis (TSA) to look at these numbers. Think of TSA as a specialized mechanic. You hand the mechanic a specific tool (like a wrench for "forecasting" or a screwdriver for "detecting anomalies"), and they fix that one specific part. They give you a number or a label (like "Normal" or "Abnormal"), but they don't really talk to you about why or what it feels like.

Recently, Large Language Models (LLMs) (like the AI behind chatbots) have entered the scene. These are like eloquent storytellers. They are amazing at reading, writing, and answering questions, but they were trained on books and conversations, not on raw numbers.

The problem? The mechanic (TSA) and the storyteller (LLM) speak different languages. The mechanic speaks "numbers," and the storyteller speaks "words."

This paper argues that we are moving from the era of the Specialized Mechanic to the era of the Conversational Guide. This new era is called Time Series Question Answering (TSQA). Instead of asking the AI to "predict the next 48 hours," you can now ask, "Why did the temperature spike yesterday, and what does that mean for my energy bill?" The AI doesn't just spit out a number; it explains the story behind the data.

The Three Ways to Connect the Mechanic and the Storyteller

The paper proposes a "menu" of three ways to make these two worlds work together, depending on how much effort and money you want to spend. They call these Alignment Paradigms.

1. Injective Alignment: The "Translator" (No Surgery)

The Analogy: Imagine you have a strict storyteller who refuses to change their voice. You take the raw numbers, write them down as a list of words (e.g., "10, 12, 15..."), and paste them into a prompt like a letter. You then say, "Here is the list, please tell me what happens next."
How it works: You don't change the AI at all. You just "inject" the numbers into its text input.
Pros: It's cheap and easy. You don't need to retrain the AI.
Cons: The AI might get confused if the numbers are messy or if the list is too long. It's like reading a grocery list and trying to guess the recipe.

2. Bridging Alignment: The "Bridge Builder" (Adding a Connector)

The Analogy: The storyteller still won't change their voice, but you build a small, custom bridge between the numbers and the words. You use a special adapter that translates the numbers into a "language" the storyteller understands better before they read it.
How it works: You add a small layer of software that converts the numbers into text-like patterns, but you still keep the main AI frozen (unchanged).
Pros: It's a great middle ground. It's more accurate than just pasting numbers, but you don't have to retrain the whole AI.
Cons: You have to build and maintain this bridge, and it might not be perfect for every type of data.

3. Internal Alignment: The "Surgery" (Changing the Brain)

The Analogy: This is the most intense approach. You take the storyteller and perform surgery on their brain. You teach them that numbers are a new language they need to learn from scratch. You might even give them new "eyes" to see charts and graphs directly.
How it works: You actually retrain the AI, updating its internal weights so it naturally understands time series, visual plots, and tables.
Pros: This creates the most powerful and flexible AI. It can "think" about time and numbers natively, not just as a translation.
Cons: It is very expensive, requires massive computing power, and takes a lot of time to train.

What the Paper Found (The Map)

The authors looked at dozens of recent research papers and sorted them into these three buckets.

Most papers right now are using the Bridge Builder approach (Bridging Alignment). It's the sweet spot for most people.
The Trend: The field is shifting. In the past, everyone was just doing specific math tasks (TSA). Now, more research is moving toward the Conversational Guide (TSQA), where the goal is to answer complex user questions rather than just calculating a number.
The Data: They analyzed datasets from medicine (heartbeats), finance (stock prices), and IoT (sensors). They found that while we have data, we often lack the "reasoning" data. We have the "what" (the numbers), but not enough "why" (the step-by-step logic) to teach the AI how to think deeply.

The Hurdles Ahead

The paper points out three main challenges we need to overcome:

Data Quality: We need better datasets that don't just give answers, but show the steps the AI should take to get there (like showing the math work on a test).
Tooling: We need to teach these AIs to use specific tools (like a calculator or a graph slicer) when they get stuck, rather than guessing.
Hallucinations: Sometimes, the AI will confidently make up a number or a trend that isn't true. Since numbers are precise, making them up is a big problem. We need better ways to stop the AI from "lying" about the data.

Summary

This paper is a roadmap. It tells us that we are moving from rigid, expert-only number crunching to flexible, user-friendly conversations with data. It offers three paths to get there (Inject, Bridge, or Surgery) and warns us that while the technology is exciting, we still need better data and tools to make the AI truly reliable.

Technical Summary: From Time Series Analysis to Question Answering: A Survey in the LLM Era

1. Problem Statement

The integration of Large Language Models (LLMs) into Time Series Analysis (TSA) has introduced a novel paradigm, leveraging natural language capabilities for tasks like forecasting and anomaly detection. However, a fundamental gap exists between traditional TSA and LLMs. Conventional TSA relies on expert-driven, task-specific pipelines (e.g., fixed forecasting or classification objectives) and predefined models. In contrast, LLMs are pre-trained to optimize natural language relevance for question answering, not for specialized TSA objectives.

Consequently, current TSA methods cannot adequately cover "temporal language tasks" such as deep interpretation, captioning, and flexible user-driven inquiry. The field is evolving from Time Series Analysis (TSA) toward Time Series Question Answering (TSQA). TSQA shifts the focus from expert-driven, task-specific analysis to user-driven, task-unified question answering. This evolution requires bridging the gap between numerical time series data and the linguistic reasoning capabilities of LLMs through effective alignment.

2. Methodology: A Taxonomy of Alignment Paradigms

The paper proposes a comprehensive taxonomy that categorizes existing literature into three distinct alignment paradigms, driven by a shift from external to internal alignment. These paradigms are defined by two dimensions: whether temporal modification (adapting time series for LLMs) is required, and whether LLM parameters are trained.

2.1 Injective Alignment

Definition: This paradigm injects numerical time series directly into textual prompts without modifying the time series data itself or the LLM architecture. It utilizes frozen LLMs.
Mechanism: Numerical sequences are encoded into token embeddings or converted into textual time series ( $X_{Text}$ ) and appended to prompts.
Optimization Strategies:
- In-Context Learning (ICL): Guides the LLM using task demonstrations (e.g., LVICL, LLMAD).
- Chain of Thought (CoT): Decomposes complex tasks into ordered intermediate steps to enhance reasoning (e.g., LSTPrompt, TimerBed).
Characteristics: Flexible and operates outside the LLM; suitable for domain-specific applications where prompt quality is critical.

2.2 Bridging Alignment

Definition: This paradigm establishes mappings between numerical time series and textual prompts, introducing temporal modification while still employing frozen LLMs.
Mechanism: A bridging module (e.g., adapters) projects time series into an embedding space compatible with the LLM.
Optimization Strategies:
- Similarity Retrieval (SR): Aligns numerical and textual representations via contrastive loss or reprogramming spaces (e.g., Time-LLM, Time-CMA).
- Low-Rank Adaptation (LoRA): Introduces parameter-efficient low-rank updates to pre-trained LLMs for task-specific tuning without altering base parameters (e.g., HiTime, AutoTimes).
Characteristics: Economical operation outside the core LLM; effective at capturing time-series characteristics through aligned semantic information.

2.3 Internal Alignment

Definition: This paradigm modifies internal components and parameters of the LLM, combining temporal modification with training the LLM.
Mechanism: Enhances time series representations (numerical, visual, or structured tables) and updates LLM weights to provide native support for time series.
Optimization Strategies:
- Re-Training (RT): Unifies temporal and textual modalities through continual pre-training and instruction tuning (e.g., ChatTime, ChatTS).
- Reinforcement Learning (RL): Aligns LLM behavior with temporal objectives using reward signals derived from sequential outcomes (e.g., TimeSense, Alpha-R1).
Characteristics: Provides generalizable, deep internal alignment; requires substantial training resources but offers superior adaptability for complex reasoning.

3. Key Contributions

Conceptual Distinction: The paper formally distinguishes and defines TSA (predefined, expert-driven) and TSQA (unified, user-driven, flexible exploration), framing the latter as the evolution of the former.
Unified Taxonomy: It proposes a taxonomy of three alignment paradigms (Injective, Bridging, Internal) that covers the evolution from TSA to TSQA. This integrates previously isolated research into a unified framework based on alignment strategies.
Practical Guidance: The authors provide actionable guidance for selecting alignment paradigms based on constraints such as computational cost, data availability, and the need for generalizability.
Dataset Analysis: The survey analyzes representative datasets across domains (Medical, Financial, IoT, General) and characteristics (temporal patterns, spatial multivariates, multimodal representations), highlighting the trade-offs between numerical, textual, and visual representations.
Challenge Identification: The paper identifies critical challenges, including the lack of reasoning supervision in datasets, the shift from explicit to implicit task extraction, and the issue of numerical hallucination in LLMs.

4. Results and Observations

The survey synthesizes 50 relevant works and presents the following distributional observations:

Paradigm Distribution: Bridging Alignment dominates the literature (50%), followed by Injective Alignment (26%) and Internal Alignment (24%).
Task Distribution: Traditional TSA tasks account for 60% of the literature, while TSQA tasks account for 40%.
Domain Coverage: The General domain (evaluated across multiple domains without specific focus) represents 74% of the literature, with Medical (10%), Financial (8%), and IoT (8%) following.
Optimization Strategies: Similarity Retrieval (SR) is the most common strategy (36%), followed by Direct injection, LoRA, and others.
Evolutionary Trends: From 2024 to 2025, the growth rate of TSQA literature has surpassed TSA. Furthermore, Internal Alignment exhibits the most significant increase in research activity, indicating a trend toward more integrated, trainable solutions.

5. Significance and Future Directions

The paper claims significance in providing the first survey that views TSA and TSQA not as isolated fields but as an evolutionary continuum driven by alignment paradigms. It argues that the shift from external alignment (Injective/Bridging) to internal alignment is essential for achieving flexible, user-driven temporal reasoning.

Future Directions Identified:

Data: Developing large-scale, multi-domain TSQA datasets that include explicit reasoning traces and diverse temporal behaviors to overcome the lack of reasoning supervision.
Task: Moving from explicit task specification to implicit extraction of user intent, and integrating dedicated temporal tools (e.g., time series slicing) into the TSQA framework.
Method: Addressing numerical hallucination and limited adaptivity. Future work should explore multi-agent mechanisms for planning and self-reflection, and investigate how alignment paradigms can better mitigate errors in continuous value prediction.

The authors conclude that while TSA provides effective task-specific solutions, TSQA offers a more intuitive, unified paradigm for temporal understanding, provided that alignment challenges regarding data quality, task definition, and model robustness are addressed.