This article didn't really clarify for me how prompt tuning works. Some examples would help. The reference to IBM mentions one key difference from prompt engineering: the use of soft prompts, which are not human-readable. That should be at the top of the Difference table.