Ribbon: New method quantifies AI uncertainty at scale without retraining

Ribbon tackles a core challenge in modern machine learning: reliably quantifying predictive uncertainty for complex, high-dimensional, or misspecified models. Traditional methods like full Bayesian inference or bootstrap resampling require expensive posterior sampling or repeated model refitting, making them impractical for large-scale models. Ribbon introduces a scalable approximation to Dirichlet-reweighted bootstrap uncertainty that replaces repeated refitting with an influence-function linearization around a single fitted model. This approach preserves the first-order data-reweighting structure of the Bayesian bootstrap while requiring only post-hoc linear algebra, dramatically reducing computational cost.

Ribbon is shown to be asymptotically equivalent to a flat-prior Laplace approximation under correct likelihood specification and recovers the robust sandwich covariance under model misspecification. The method includes a general concentration parameter that allows tuning of uncertainty scale on validation data. Across synthetic regression, MNIST classification, and California Housing benchmarks, Ribbon provides competitive predictive performance and improved calibration in several settings while avoiding repeated model training. This makes it a practical tool for practitioners who need trustworthy uncertainty estimates for high-stakes decisions.