I wonder too. I imagine the best we could do with present technology is to get back the generated text in the form of text tokens accompanied by their corresponding deep embeddings (last hidden states): `[(text_token, deep_emb), (text_token, deep_emb), ...]`. Those deep embeddings incorporate "everything the model knows" about each generated token of text.
Maybe a mapping/representation for a "medium embedding" could be learned that strikes a balance between shallow and deep. I have no idea what a good objective-function would be, though.
