WAG beats RAG by 70% for LLM reasoning on wearable data
New graph-based retrieval method boosts LLM performance on 10,000 wearable queries.
Large language models (LLMs) are increasingly applied to wearable health data, but traditional approaches struggle with context selection—either providing too little information for accurate reasoning or overwhelming the model with irrelevant data. Researchers from the University of California, Irvine, propose Wearable As Graph (WAG), a graph-based retrieval framework that dynamically constructs a personalized knowledge graph from wearable metrics and user-specific signals. When a query arrives, WAG retrieves a query-conditioned subgraph that balances global relationships (captured via hierarchical Bayesian modeling of population and individual patterns) with local relationships (short-term signal deviations).
WAG was evaluated on over 10,000 data-grounded queries from real-world wearable datasets, comparing against standard RAG methods using both automated LLM-based and human evaluations. The framework achieved an approximately 70% win rate across all metrics, demonstrating that structured, query-adaptive context retrieval significantly improves LLM reasoning on personalized, multimodal, time-series data. This approach could unlock more accurate health insights, anomaly detection, and behavioral predictions from continuous wearable monitoring without sacrificing efficiency.
- WAG organizes wearable data into a personalized knowledge graph, using hierarchical Bayesian modeling to capture both global patterns and local signal deviations.
- Achieves ~70% win rate over standard RAG methods across over 10,000 real-world wearable queries.
- A query openness signal dynamically controls retrieval breadth, balancing context sufficiency with efficiency.
Why It Matters
Enables LLMs to reason accurately on continuous wearable health data, improving personalized diagnostics and activity insights.