Agentic Sensor Language Models for Reasoning over Time-Series

Recent efforts applying LLMs for time-series analysis and classification tasks, typically by converting time-series signals to text or images to be processed by LLMs or VLMs, suffer from poor performance and lack of interpretability. We have pushed the boundaries of using LLMs for time-series classification through SensorLLM and ZARA. In Sensor LLM, we introduced a two-stage framework—first aligning sensor data with intuitive trend descriptions, then fine-tuning for classification—achieving accuracy on par with or surpassing state-of-the-art supervised models. In contrast, ZARA eliminates retraining altogether, proposing an agent-based, zero-shot framework that combines a feature knowledge base, evidence retrieval, and hierarchical reasoning to predict activities while also providing natural-language explanations. While SensorLLM highlights generalization through human-intuitive alignment and supervised finetuning, ZARA emphasizes interpretability and plug-and-play adaptability, achieving strong zero-shot performance and explainability.

This Taste of Research project will build on our past research. The focus of the project is on verifiable reasoning, since most current reasoning/thinking lacks explicit evidence support. The overarching goal is to enhance time-series reasoning capabilities over a broad time-series tasks, enabling robust QA and chat-based exploration over high dimensional time-series.