Abstract

Recognizing human activities from IMU sensor data underpins health monitoring, smart environments and HCI. Conventional HAR methods hinge on complex feature engineering or deep learning, demanding extensive labels and compute. We investigate repurposing LLMs for IMU‑based HAR by encoding statistical and spectral features into text prompts, enabling zero‑ and few‑shot classification. We also explore hybrid models where LLMs enrich classical classifiers with metadata and high‑level semantics. Preliminary experiments show LLMs rival traditional methods in interpretability and adaptability—especially in low‑data or personalized scenarios—paving the way for more transparent, generalizable and interactive HAR systems.

Research in progress