**Title: Gravity-Aware Hierarchical Routing for Lightweight SensorLLM on Human Activity Recognition

Abstract:

While recent research into sensor-language alignment suggests that two-stage frameworks enhance the semantic modeling capabilities of wearable-sensor human activity recognition (HAR)—specifically through SensorLLM-style approaches that first align motion to language before fine-tuning for downstream tasks—our experiments uncover a significant limitation. When the Stage 2 backbone is compressed into a compact architecture like TinyLlama, the model exhibits a consistent failure mode: although the recognition of dynamic activities remains robust, the ability to distinguish low-motion static classes, such as standing, sitting, and lying, deteriorates significantly.

To mitigate this issue, we introduce a gravity-aware hierarchical routing head. This approach serves as a lightweight post-alignment adaptation for an already aligned model, eliminating the need for a new large-scale pretraining framework. By leveraging per-channel mean and standard deviation from the Chronos tokenizer state, the method extracts statistical indicators related to posture and gravity direction. It then adaptively merges a static expert with a full expert via soft routing, supported by a load-balancing loss to ensure training stability.

Evaluated on the MHealth dataset, this design yields substantial improvements in macro-F1 with negligible parameter overhead. The performance gains are primarily driven by enhanced accuracy on static classes, while maintaining strong results for dynamic activities. As this is the first arXiv disclosure of this work, the current report focuses on a single dataset to highlight the core methodology, laying the foundation for more extensive evaluations in subsequent studies.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC