📑 Report
The increasing prevalence of time-series data across domains, such as healthcare, finance, and IoT, has driven the need for flexible and generalizable modeling approaches. Traditional time-series analysis relies heavily on supervised models, which suffer from data scarcity and high deployment costs. This project explores the potential of Large Language Models (LLMs) as general-purpose few-shot learners for time-series classification tasks. Specifically, we investigate how different input representations affect LLM performance in two representative tasks: Dual-Tone Multi-Frequency (DTMF) signal decoding and Human Activity Recognition (HAR). By comparing text-based, visual, and multimodal inputs across models including GPT-4o, GPT-o1, and DeepSeek-R1, we assess their reasoning capabilities and robustness. Our findings show that while LLMs demonstrate potential, performance significantly varies depending on input representation and model type. In DTMF, GPT-o1 and DeepSeek-R1 consistently outperforms GPT-4o, particularly in tasks requiring text-based numerical reasoning. In HAR, visualization aids interpretation, and few-shot learning significantly boosts performance. However, challenges remain, especially in precise plot reading, domain knowledge retrieval (GPT-4o), and multimodal integration. Further domain-specific enhancements and robust representation strategies are heavily required for current LLMs. Code for this project is available on https://github.com/nesl/LLM_time_series.