Clustered Randomized Smoothing for Stochastic Prediction Functions

What the math gives to ML

The paper's transferable contribution is a mode-preserving alternative to global randomized smoothing for stochastic predictors. Instead of averaging all noisy predictions, it partitions prediction samples into clusters and applies coordinate-wise alpha-trimming within each cluster, producing a mixture of locally robust components. This is directly useful as an inference-time wrapper for trajectory, diffusion, policy, or multimodal regression networks, where global smoothing can average incompatible modes into unsafe outputs. The strongest initial implementation is a black-box clustered smoothing layer that returns cluster weights, trimmed centers, and within-cluster covariance, followed by experiments measuring distributional fidelity and safety under input perturbations.