Training-Free Time Series Classification via In-Context Reasoning with LLM Agents

1. Problem Statement

Time Series Classification (TSC) is a fundamental task in domains like healthcare, finance, and industrial monitoring. However, existing solutions face significant limitations:

• Data Scarcity & Cost: Supervised deep learning models require large amounts of labeled data and expensive, dataset-specific training or fine-tuning.
• Inflexibility: Traditional models (e.g., CNNs, Transformers) and even recent LLM-based approaches often rely on pretraining or fine-tuning, making them difficult to deploy in "plug-and-play" scenarios where labeled data is scarce.
• Suboptimal Zero-Shot Performance: While reasoning-oriented Large Language Models (LLMs) show promise in understanding temporal patterns (trends, seasonality), using them in a purely zero-shot or fixed-prompt manner yields suboptimal results.

The Core Question: Can we perform multivariate time series classification effectively without any model training, pretraining, or fine-tuning, leveraging the in-context learning capabilities of modern reasoning LLMs?

2. Methodology: The FETA Framework

The authors propose FETA (training-Free timE series classificaTion with LLM Agents), a multi-agent framework that replaces model fitting with exemplar-based in-context reasoning. The pipeline consists of four specialized agents operating in a modular workflow:

A. Channel Decomposer (Selection & Preprocessing)

Multivariate time series often contain redundant or irrelevant channels and varying sequence lengths, which hinder LLM reasoning efficiency.

• Function: Decomposes the multivariate input into univariate channels and prunes irrelevant ones.
• Mechanism:
  1. Normalization: Applies length normalization and per-sample z-normalization to ensure comparability.
  2. Scoring: Computes a training-free relevance score for each channel based on two metrics:
     • Prototype-Margin Score ($B_c$): Measures inter-class separation vs. intra-class compactness (similar to Fisher's ratio).
     • Approximate 1NN Leave-One-Out Accuracy ($C_c$): Evaluates standalone classification ability using nearest neighbors.
  3. Selection: Fuses these scores to rank channels and selects the top-$M$ most informative channels for downstream processing.

B. Example Retriever (In-Context Exemplar Selection)

To ground the LLM's reasoning, the system retrieves relevant labeled examples for the query.

• Function: Retrieves structurally similar labeled sequences from the training set for each selected channel.
• Mechanism: Uses Dynamic Time Warping (DTW) to find the $K_r$ nearest neighbors. DTW is chosen over Euclidean distance because it robustly handles temporal distortions, variable speeds, and phase shifts.
• Output: A set of exemplars $\{(X_i, y_i, d_i)\}$ containing the sequence, label, and alignment-based distance.

C. Channel Reasoner (Pattern-Based In-Context Reasoning)

This is the core reasoning engine where the LLM operates.

• Function: Performs in-context comparison of the query against retrieved exemplars to generate a channel-level prediction and confidence.
• Mechanism:
  • Each selected channel is assigned an independent agent.
  • The agent receives the query sequence and the retrieved exemplars (with DTW distances).
  • The reasoning LLM analyzes temporal patterns (trends, fluctuations, turning points) to adjudicate between conflicting exemplar labels.
  • Output: A predicted label ($\hat{y}_c$) and a self-assessed confidence score ($w_c \in [0, 1]$).
  • Key Advantage: This step moves beyond simple majority voting by leveraging the LLM's ability to interpret complex temporal nuances.

D. Decision Aggregator (Confidence-Weighted Fusion)

The final stage synthesizes decisions from all channel agents.

• Function: Aggregates channel-level predictions into a single global decision.
• Mechanism:
  • Consensus Mode: If all agents agree, the confidence is boosted based on the number of agreeing agents.
  • Weighted Mode: If agents disagree, a confidence-weighted fusion is applied. Valid predictions contribute their confidence scores to the predicted class, while a small smoothing weight is added to other classes.
  • Result: The class with the highest aggregated score is selected. This approach is more robust than naive majority voting, especially when channels provide conflicting evidence.

3. Key Contributions

1. First Training-Free Multi-Agent Framework: Introduces the first framework for multivariate TSC that leverages LLM in-context reasoning without any parameter updates, pretraining, or fine-tuning.
2. Modular Pipeline Design: Proposes a novel architecture integrating channel decomposition, DTW-based exemplar retrieval, LLM reasoning, and confidence-weighted aggregation. This ensures efficiency (by pruning channels) and interpretability (via exemplar grounding).
3. Empirical Validation: Demonstrates that this approach achieves state-of-the-art (SOTA) performance on challenging benchmarks, outperforming both classical methods and trained deep learning models.

4. Experimental Results

• Datasets: Evaluated on nine representative and challenging datasets from the UEA Time Series Classification Archive.
• Baselines: Compared against distance-based methods (DTW), pattern-based methods (WEASEL+MUSE), deep learning models (MLSTM-FCNs, TST), representation learning methods (TS-TCC, TS2Vec), and other multi-agent approaches (MERIT).
• Performance:
  • FETA (Qwen3) achieved the highest average accuracy of 47.3%, surpassing the best trained baseline (MERIT at 44.4%) and classical methods.
  • FETA (DeepSeek R1) achieved 46.6%, ranking second overall.
  • Even FETA (LLaMA 3.1), using a smaller non-reasoning model, achieved 38.9%, matching the Transformer-based TST baseline and significantly outperforming a direct zero-shot DeepSeek R1 baseline (32.5%).
• Ablation Studies: Confirmed that all components are critical. Removing the Channel Decomposer caused the largest performance drop (–14.6% average), highlighting the importance of input optimization. Removing the Reasoner or Aggregator also significantly degraded performance, proving the necessity of LLM reasoning and confidence-weighted fusion over simple heuristics.

5. Significance and Impact

• Paradigm Shift: FETA demonstrates that LLMs can serve as competitive, plug-and-play solvers for time series tasks without the computational cost and data requirements of traditional training pipelines.
• Interpretability: Unlike "black box" deep learning models, FETA provides interpretable decisions grounded in specific retrieved exemplars and explicit confidence scores.
• Efficiency: By pruning irrelevant channels and controlling input length, the framework makes LLM reasoning feasible for high-dimensional time series data.
• Future Direction: This work opens a new avenue for integrating reasoning LLMs with structured temporal data under zero-training constraints, suggesting that "learning" in time series can be achieved through retrieval and reasoning rather than parameter optimization.

Limitations: Currently designed for structured numerical signals; future work needs to address hybrid data modalities (e.g., text, images, metadata) and the environmental cost of large-scale LLM inference.