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

Hao Li; Mingrui Zheng; Yasuyuki Tahara; Yuichi Sei

← Publications

Preprint 2026

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

Hao Li, Mingrui Zheng, Yasuyuki Tahara and Yuichi Sei^*

arXiv preprint, arXiv:2606.04019 (2026)

arXiv ↗

ABSTRACT

Abstract

This preprint examines a failure mode of compact SensorLLM-style models in which dynamic activities remain recognizable while low-motion static postures become difficult to distinguish. It introduces a lightweight gravity-aware hierarchical routing head that uses statistics from the tokenizer state to route between static and full experts. On the MHealth dataset, the method improves macro-F1 mainly on static classes while retaining strong performance on dynamic activities with little parameter overhead.

CITATION

BibTeX

@article{Li2026GravityAware,
  author  = {Hao Li and Mingrui Zheng and Yasuyuki Tahara and Yuichi Sei},
  title   = {Gravity-Aware Hierarchical Routing for Lightweight SensorLLM on Human Activity Recognition},
  journal = {arXiv preprint arXiv:2606.04019},
  year    = {2026},
  doi     = {10.48550/arXiv.2606.04019},
  url     = {https://arxiv.org/abs/2606.04019}
}