
Improving Generative AI Student Feedback: Direct Preference Optimization with Teachers in the Loop
Juliette Woodrow, Chris Piech, Sanmi Koyejo
Proceedings of the 18th International Conference on Educational Data Mining
Teaching assistants' choices can help adapt AI feedback to a particular course. In a university probability course, researchers trained a model using comments that staff selected, edited, or wrote themselves. Experts preferred its feedback to GPT-4o in a blind comparison, while classroom staff still frequently wrote their own responses. The study demonstrates a feasible feedback workflow and stronger teacher preference, but it did not test whether students learned more or used the feedback more effectively.
