Learned embeddings or activation patterns that capture the essence of an alignment value for prediction tasks.