Adaptive Sharpness Regularization for Probabilistic Circuits

A new compositional theory of curvature shows that global sharpness regularization in Probabilistic Circuits (PCs) is often misspecified. The authors introduce an adaptive regularizer that penalizes nodes based on their local curvature and circuit flow to prevent depth-biased underfitting.