How do predictions shift when environments change?

Real-world prediction often fails when data comes from a new hospital, lab, or observational setting. This work assumes that hidden factors change across environments (e.g., patient prevalence), but the core relationships between measurements and outcomes stay fixed. The authors build a Bayesian model that learns these stable latent structures and a variational algorithm to infer them, then use those learned patterns to predict in new environments. Tests on disease detection, sepsis prediction, and astronomy show consistent gains over existing transfer methods.