An Exploratory Study to Repurpose LLMs to a Unified Architecture for Time Series Classification

TL;DR

This work investigates repurposing LLMs for time series classification by pairing a frozen LLM with dedicated time series encoders. By evaluating multiple encoder families (MLP, CNN, Inception, ResNet, Transformer) on the 2015 UCR Time Series Archive, the study demonstrates that Inception-based encoders uniquely provide consistent gains when integrated with an LLM, likely due to multi-scale temporal feature extraction. The findings underscore the critical role of encoder architecture in hybrid LLM-TSC systems and establish Inception-like designs as a promising direction for future LLM-driven time series analysis. Overall, the approach offers a reusable framework that leverages latent temporal representations to enable high-level semantic reasoning via LLMs, reducing reliance on domain-specific handcrafted features.

Abstract

Time series classification (TSC) is a core machine learning problem with broad applications. Recently there has been growing interest in repurposing large language models (LLMs) for TSC, motivated by their strong reasoning and generalization ability. Prior work has primarily focused on alignment strategies that explicitly map time series data into the textual domain; however, the choice of time series encoder architecture remains underexplored. In this work, we conduct an exploratory study of hybrid architectures that combine specialized time series encoders with a frozen LLM backbone. We evaluate a diverse set of encoder families, including Inception, convolutional, residual, transformer-based, and multilayer perceptron architectures, among which the Inception model is the only encoder architecture that consistently yields positive performance gains when integrated with an LLM backbone. Overall, this study highlights the impact of time series encoder choice in hybrid LLM architectures and points to Inception-based models as a promising direction for future LLM-driven time series learning.