arXiv:2607.03089v1 Announce Type: cross Abstract: HAR is increasingly expected to run continuously on edge devices, yet recent LLM-based methods remain hard to deploy: raw sensor prompts are long, cloud inference adds latency and privacy risk, and fine-tuned LLM pipelines turn general-purpose models into task-specific classifiers. We present STELLA, an efficient sensor-to-LLM translation framework for on-device HAR that shifts the burden from LLM adaptation to sensor tokenization. A lightweight hierarchical tokenizer compresses an entire multi-channel inertial window into a fixed set of compac
The proliferation of LLMs and increasing demand for real-time, privacy-preserving AI on edge devices are driving innovation in efficient on-device AI implementations.
This development allows sophisticated human activity recognition to run directly on smartphones and wearables, enabling richer AI experiences without reliance on cloud infrastructure, enhancing privacy and reducing latency.
The burden of deploying LLM-based HAR shifts from complex LLM adaptation to efficient sensor tokenization, making advanced activity recognition more practical for edge devices.
- · Edge AI chip manufacturers
- · Wearable device companies
- · Privacy-focused AI applications
- · Developers of on-device AI models
- · Cloud-dependent HAR solutions
- · Generalized LLM deployment models for specific tasks
More sophisticated, real-time human activity recognition becomes widely available on personal devices.
This could lead to new applications in health monitoring, personal security, and context-aware computing, further increasing demand for efficient on-device AI.
The enhanced privacy and data locality of such systems may accelerate the adoption of personalized AI, reducing reliance on centralized data processing for sensitive applications.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI