STELLA Enables Efficient On-Device Human Activity Recognition with LLMs.

Human Activity Recognition (HAR) is increasingly expected to operate continuously on edge devices. However, current LLM-based HAR methods face significant deployment challenges, including long raw sensor prompts, cloud inference latency and privacy risks, and the transformation of general-purpose LLMs into task-specific classifiers through fine-tuning. Researchers present STELLA, an efficient framework designed for sensor-to-LLM translation specifically for on-device HAR. STELLA shifts the computational burden from LLM adaptation to sensor tokenization. It employs a lightweight hierarchical tokenizer that compresses an entire multi-channel inertial window into a fixed set of compact latent sensor tokens. These tokens are then projected into the embedding space of a frozen, pre-trained LLM and combined with a natural-language prompt for label scoring. This approach effectively preserves activity-relevant temporal and cross-channel structure while ensuring predictable LLM-side computation across diverse sensor configurations. STELLA also supports on-device personalization by adapting only the lightweight tokenizer with small amounts of user-specific labeled data. It augments inference with a local retrieval context, keeping the LLM, user data, and retrieval entirely on the device, enhancing privacy. Across seven public HAR datasets and eight benchmark settings, STELLA achieved new state-of-the-art performance, improving over prior methods by up to 11.83% F1. On-device personalization further boosted performance by up to 21.91% F1 as user data accumulated. STELLA also outperformed representative time-series tokenizers under the same LLM pipeline and achieved real-time inference within practical mobile and edge device budgets, demonstrating that efficient sensor tokenization is a viable path for accurate, private, and personalized LLM-based HAR on edge devices.