Research papers monitoring

A preprint by Zechen Li and 13 co-authors at Google Research and collaborators (arXiv 2605.30865, submitted May 29, highlighted August 26) tackles slow states and short events mixing in continuous glucose monitoring, with labels scarce. Its contribution is a dual-stream self-supervised model trained on 109,066 hours across 477 records, evaluated on 4 cohorts and 7 tasks. The work matters for transferable metabolic-risk research. There is no independent reproduction, clinical deployment or regulatory validation, and the blog and abstract report the PR-AUC gain as 5.8 and 4.1 points under different comparator settings.

A second entry is a company technical release with an open repository, not a peer-reviewed paper. The sparse-attention, residual, embedding, optimisation and N-gram embedding claims are developer-reported.