A training loss that teaches one model to reproduce another model's ranking of candidates using Kullback-Leibler divergence.