WEQA Wearable Health QA: 24% higher accuracy than LLM baselines

WEQA is a query-adaptive agent framework for answering questions about wearable health data, submitted to arXiv on 16 Jun 2026 by Yuwei Zhang and coauthors. The system uses an LLM controller to synthesize execution plans, dynamically route queries to sensor analysis and pretrained models, and perform grounded response auditing with external knowledge.

What is WEQA and how does it work?

WEQA is a query-adaptive agent framework that pairs a language model controller with specialized wearable analytics and pretrained models, synthesizing execution plans and dynamically routing each query to the appropriate sensor analyses, then performing grounded response auditing with external knowledge to produce answers. The paper describes an LLM controller that creates execution plans and routes queries to combinations of sensor analysis tools and pretrained models tailored to the query and the modalities involved.

WEQA is designed specifically for wearable data, which the authors note are continuous, high-dimensional, and longitudinal and therefore do not align well with text-centric LLM pretraining. To handle diverse sensor modalities and user intents, WEQA does not rely on a single fixed reasoning workflow or a single pretrained foundation model; instead it composes specialized analytical and predictive tools as needed per query.

How well does WEQA perform?

In experiments on a curated benchmark spanning four open wearable datasets, WEQA is 24% more accurate than LLM and agentic baselines, and a blinded study with 12 medical experts and 8 users found substantial gains in usefulness and clinical soundness. The benchmark covers analytic and predictive tasks across three different health domains, enabling comparisons on both types of problems.

The paper positions these results against two persistent challenges: aligning continuous wearable signals with text-trained LLM distributions, and accommodating diverse sensor modalities and user intents. WEQA tackles those by routing queries to modality-appropriate analyses and by auditing outputs against external knowledge sources before returning a response.

Why it matters

Large language models can already perform well on many medical question answering tasks, sometimes matching or exceeding general physicians, but wearable health questions remain underexplored because sensor data differ dramatically from text. WEQA's architecture addresses that gap by combining LLM planning with dedicated analytical tools for wearable signals, showing measurable accuracy and clinician-perceived quality improvements in the authors' experiments.

That combination matters because wearable devices generate persistent, high-dimensional data streams; a single-text-centric LLM workflow is unlikely to reliably interpret those signals. Demonstrating a 24% accuracy improvement on a multi-dataset benchmark and positive blinded expert feedback suggests a practical path for safer, more useful wearable QA systems.

What to watch

Check the paper's arXiv entry for the authors' linked code, data, and demos: the submission page lists associated code, data, and demo toggles. The clearest next milestones will be public releases of those artifacts and larger clinical evaluations beyond the blinded study with 12 medical experts and 8 users, which would validate WEQA's gains at scale.

Authors: Yuwei Zhang, Tong Xia, Bianca Emmerich, Yu Yvonne Wu, Dimitris Spathis, Xin Liu, Daniel McDuff, Cecilia Mascolo. Submission date: 16 Jun 2026.