Exactly. It would be amazing if the same LLM were used for both generation and grading, and could see what tweaks the human (student) made before turning it in.

Exactly. It would be amazing if the same LLM were used for both generation and grading, and could see what tweaks the human (student) made before turning it in. That would be very useful training data.