Have you found any particularly promising approach towards "hypothetical embeddings" that you can recommend?