Ai2's TutorMoments evaluates whether AI tutors know when to hold back

Built from transcripts of real one-on-one math tutoring sessions, TutorMoments replays decision points where tutors chose between helping and pushing students to reason — finding LLMs systematically over-help. Prompting the trade-off improves performance but doesn't match human tutors; Ai2 released the dataset, code, and model replays.