Wearable As Graph: Query-Adaptive Context Retrieval for LLMs on Personal Data
A new framework called Wearable As Graph (WAG) has been introduced by researchers to enhance the reasoning capabilities of large language models (LLMs) by retrieving context that adapts to queries from extensive, multimodal, and personalized wearable sensing data. WAG structures wearable metrics and individual-specific signals into a tailored knowledge graph, allowing for the extraction of a subgraph conditioned on queries. This retrieval process combines global relationships through hierarchical Bayesian modeling—capturing both population and individual patterns—with local relationships that reflect short-term signal variations. A signal indicating query openness regulates the extent of retrieval. This method tackles the critical issue of context selection, balancing the need for sufficient context with the risks of inefficiency and compromised generation quality. The paper can be found on arXiv (2605.18763).
Key facts
- WAG stands for Wearable As Graph.
- The framework is designed for LLM reasoning over wearable sensing data.
- Data is long-term, multimodal, and highly personalized.
- Context selection is a key challenge addressed.
- WAG organizes data into a personalized knowledge graph.
- Retrieval is query-conditioned and integrates global and local relationships.
- Global relationships use hierarchical Bayesian modeling.
- Local relationships reflect short-term signal deviations.
- A query openness signal controls retrieval breadth.
- The paper is on arXiv with ID 2605.18763.
Entities
Institutions
- arXiv