Contrastive Linearized Laplace: Loss Geometry, Valid Repairs, and Reliable Cores
The framework uses Linearized Laplace Approximation (LLA) to propagate encoder uncertainty through downstream tasks. It moves beyond simple representation variance by analyzing the curvature of contrastive objectives and applying specific mathematical repairs to ensure valid local Gaussian approximations. This yields 'reliable cores' that quantify decision stability rather than just embedding variance.