Human activity recognition from sensors is moving beyond purely statistical models. The new framework, OntoAR-HAR, uses large language models to connect raw motion signals with higher-level activity semantics and physical knowledge. Rather than relying on the LLM alone, it introduces an ontology-guided agentic retrieval-augmented generation pipeline that retrieves relevant sensor examples and domain rules to ground each decision.
A key addition is personalized memory, which lets the system retain user-specific patterns and adapt its recognition over time. This addresses a common limitation of generic HAR models that treat all users the same. The authors position the approach as a way to make LLM-based activity recognition more robust and context-aware, though the abstract does not report quantitative results.
The paper is available on arXiv under the identifier 2610.15292. As a single-source digest, this summary reflects only the claims in the abstract; no independent evaluation or comparison with other methods is included here.