Comments

However, existing approaches only use the “direction” of the unitary embedding vectors.

This is the biggest crux for me. It sounds like this project is largely a bet that magnitude is also important. Do you have evidence for that?