Meta researchers have created a novel and very useful health benchmark: WearableQA As LLMs are starting to be used to analyze longituditinal wearable data for personalized medicine applications, it is important to systematically benchmark this capability on real-world data. Here, the authors introduce WearableQA, a benchmark for data and health reasoning on real-user longitudinal wearable records. WearableQA contains 4,084 multiple-choice questions constructed from 200 real users’ wearable measurements (16 daily metrics) collected over hundreds of days per individual, with complementary blood biomarkers (17 biomarker blood panel) and demographic information. The benchmark has two axes of questions: * reasoning type: some questions require data reasoning (computing trends, anomalies, and relationships from the raw wearable measurements) while others require health reasoning (clinical and physiological interpretation) * signal complexity: there are single-signal questions vs. cross-signal questions The authors evaluated a total of XXX models, both proprietary and open-source. A few conclusions: * performance ranges widely, from 72.9% accuracy to 19.6% against a 10% chance baseline. * Data reasoning remains more challenging than health reasoning for most models. * Cross-signal reasoning remains challenging across model scales. Note that the benchmark hasn't been released yet, though the paper has links for HuggingFace and GitHub, so the release is likely soon. Paper link: arxiv.org/abs/2609.05405