TLDR: FETA is a novel multi-agent framework that enables training-free time series classification (TSC) by leveraging large language models (LLMs) for in-context reasoning. It decomposes multivariate time series into channel-wise subproblems, retrieves structurally similar labeled examples using Dynamic Time Warping (DTW), uses LLMs to compare queries against these exemplars and produce channel-level labels with confidence, and finally fuses these decisions using a confidence-weighted aggregator. This approach eliminates the need for pretraining or fine-tuning, improves efficiency, and enhances interpretability, achieving strong accuracy on challenging datasets, often surpassing traditional trained baselines.
Time series classification (TSC) is a critical task in many fields, from healthcare to finance and industrial monitoring. However, a major challenge is the scarcity of labeled data, which makes traditional training methods for these classification models expensive and inflexible. This often requires significant computational power, hyperparameter tuning, and effort to adapt models to new domains.
Recent advancements in large language models (LLMs) have shown a remarkable ability to understand complex patterns, including temporal ones. While these models hold great promise for time series analysis, simply using them in a “zero-shot” manner (without any specific training for the task) often doesn’t yield optimal results. Many existing LLM applications for time series still rely on additional training or converting time series data into text-like tokens, which limits their versatility and ease of use.
Addressing this gap, researchers Songyuan Sui, Zihang Xu, Yu-Neng Chuang, Kwei-Herng Lai, and Xia Hu have introduced a novel framework called FETA (training-Free timE series classificaTion with LLM Agents). FETA proposes a completely training-free approach to time series classification by leveraging exemplar-based in-context reasoning. This means the system learns from examples provided in the moment, rather than through a lengthy pre-training or fine-tuning process.
How FETA Works: A Multi-Agent Approach
FETA operates as a multi-agent framework, where different specialized “agents” collaborate to classify a time series. This modular design eliminates the need for traditional model training, improves efficiency by focusing on relevant information, and enhances interpretability by showing how decisions are made based on specific examples.
The framework consists of four key agents:
Channel Decomposer: This agent is the first step. It takes a complex multivariate time series (data with multiple simultaneous measurements, or “channels”) and breaks it down into individual, simpler channel-wise subproblems. It also intelligently filters out less relevant channels, ensuring that the subsequent reasoning focuses only on the most informative signals. This process involves normalizing data and ranking channels based on how well they separate different classes and support classification.
Example Retriever: Once the informative channels are identified, this agent steps in. For each channel, it searches through a collection of labeled training data to find a few structurally similar examples. It uses a technique called Dynamic Time Warping (DTW), which is excellent at finding similarities between time series even if they have slight temporal shifts or different speeds. These retrieved examples serve as “in-context” demonstrations, guiding the LLM’s reasoning without any parameter updates.
Channel Reasoner: This is where the LLM’s reasoning power comes into play. For each channel, a dedicated agent (powered by a reasoning LLM) compares the query time series with the examples retrieved by the Example Retriever. It analyzes temporal patterns like trends, fluctuations, and alignments. Instead of just voting on a label, the LLM provides a predicted label for that channel along with a self-assessed confidence score, explaining its decision based on the provided exemplars.
Decision Aggregator: The final agent collects all the channel-level predictions and their confidence scores. Since different channels might offer complementary or even conflicting evidence, the aggregator fuses these decisions into a single, robust final classification. It prioritizes agreement among agents with high confidence and uses a confidence-weighted approach when predictions are not unanimous, ensuring a balanced and interpretable outcome.
Also Read:
- SPEAR: Enhancing Anomaly Detection in Time Series Data with Soft Prompts and Language Models
- ContextNav: An Agentic Approach to Better Multimodal In-Context Learning
Performance and Impact
FETA has been rigorously tested on nine challenging UEA datasets, which are standard benchmarks for time series classification. The results demonstrate that FETA consistently achieves strong accuracy, often surpassing multiple traditional trained baselines, including deep learning and representation learning models. For instance, FETA using Qwen3 achieved an average accuracy of 47.3%, outperforming many established methods.
The success of FETA is attributed to its innovative design: the effective decomposition and selection of channels, the use of DTW-based exemplar retrieval for grounding reasoning, the confidence-aware fusion of decisions, and the overall modular multi-agent architecture. This framework transforms LLMs into competitive, plug-and-play time series classification solvers without requiring any parameter training.
This research opens up new possibilities for integrating reasoning LLMs with structured temporal data, especially in scenarios where labeled data is scarce and training new models is impractical. The code for FETA is publicly available, fostering further research and application in this exciting area. You can find more details about the FETA framework in the full research paper: Training-Free Time Series Classification via In-Context Reasoning with LLM Agents.
