A new approach, Loki, adapts pretrained models to predict new classes without additional training by using a metric that relates labels via distances. This can improve model performance on zero-shot prediction tasks.