Researchers have a plan to make your smartwatch's sleep coach smarter without shipping your data to a cloud AI.
The team built a two-stage system for turning raw wearable sleep data into personalized guidance. First, a multi-agent large language model pipeline reads unannotated wearable records and reasons out structured sleep advice, effectively building a training dataset without human annotators. Second, that reasoning gets distilled into small language models through supervised fine-tuning, plus a training-free "best-of-N" selection step that picks the strongest of several generated answers at inference time. The researchers report their distilled small models beat both general-purpose commercial LLMs and specialized medical LLMs, as well as open-source alternatives, on the sleep-guidance task, with human evaluators rating the output as good quality.
Why it matters: expert-annotated health data is expensive and slow to collect, which is why most sleep-guidance AI stays generic or never ships at all. Distilling reasoning from large models into small ones that run locally sidesteps both the annotation bottleneck and the privacy problem of sending sleep data to a third-party server.
It's a clever workaround, but "outperforms commercial LLMs" on a benchmark the researchers built themselves is a claim worth revisiting once someone tests it on data they didn't curate.