STELLA introduces a sensor-to-LLM framework for on-device human activity recognition. Its lightweight hierarchical tokenizer compresses a multi-channel inertial window into a fixed number of latent sensor tokens, projects them into the embedding space of a frozen pretrained LLM, and combines them with a natural-language prompt for label scoring. The paper reports results across seven public HAR datasets and eight benchmark settings, with up to 11.83% F1 improvement over prior methods. Personalizing only the tokenizer with small user-labeled datasets reportedly adds up to 21.91% F1 as data accumulates, while keeping inference, retrieval, and user data on device.
No heat snapshots are available in the last 24 hours.