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

Researchers propose a gravity-aware hierarchical routing method to improve human activity recognition in compressed language models used with wearable sensors. The lightweight adaptation addresses a specific failure mode where static activities like standing and sitting are poorly recognized when using compact models like TinyLlama, while maintaining strong performance on dynamic activities.

This research addresses a practical limitation in deploying sensor-language models on resource-constrained wearable devices. The two-stage SensorLLM framework—which aligns motion data with language representations before fine-tuning—has shown promise for activity recognition, but compressing the backbone model introduces a critical trade-off: dynamic activity recognition remains robust while static postures degrade significantly. This asymmetric performance degradation suggests that compressed models lose fine-grained discriminative capacity for subtle postural distinctions. The proposed gravity-aware hierarchical routing head tackles this by leveraging statistical features from the tokenizer state to route inputs adaptively between a specialized static expert and a full expert network. By extracting gravity-direction cues from channel-wise statistics, the method achieves targeted improvements for low-motion classes without requiring large-scale retraining. The approach demonstrates practical engineering: rather than rebuilding pretraining frameworks, the solution operates as a lightweight post-alignment adapter, adding minimal parameters while improving macro-F1 scores on the MHealth dataset. This design philosophy aligns with industry trends toward efficient edge AI, where practitioners balance model compression against task-specific performance. However, the current work's limitation to a single dataset constrains confidence in generalization across diverse activity types and sensor modalities. The research lays groundwork for investigating whether similar routing strategies could address compression-induced performance asymmetries in other sensor-language applications, though broader validation across multiple datasets and hardware platforms remains necessary before widespread deployment consideration.

• →Compressed sensor-language models exhibit selective performance degradation, losing discrimination on static activities while maintaining dynamic activity recognition.
• →Gravity-aware routing leverages tokenizer statistics to adaptively specialize processing for static versus dynamic postures with minimal parameter overhead.
• →The hierarchical expert routing approach achieves gains concentrated on low-motion classes without retraining underlying aligned models.
• →Single-dataset validation limits generalization claims; broader evaluation across datasets and sensor types is needed for production deployment.
• →Lightweight post-alignment adapters offer a practical alternative to large-scale model retraining for addressing specific performance bottlenecks.