What is KL divergence used for in model evaluation and monitoring?

KL divergence measures how one probability distribution differs from another.

KL divergence quantifies information loss when approximating distribution P with Q. In model evaluation, it is used to detect drift, compare predicted vs true distributions, and monitor changes in feature distributions. Unlike symmetric metrics, KL is directional and penalizes underestimation heavily. It is widely used in probabilistic modeling, VAEs, and monitoring pipelines.