AI2's TutorMoments finds helpfulness training fights any task where withholding is correct.

TutorMoments replays real decision points from 462 de-identified one-on-one math tutoring transcripts with 1,500+ teacher-marked moments and thousands of annotations from 27 educators. Across seven models, default assistant behavior was to over-scaffold, making problems easier instead of letting students reason. Performance improved substantially once the prompt explicitly named the scaffolding-versus-rigor trade-off, though variance stayed wide. Transferable well beyond education: if your task requires the model to hold back, you have to say so explicitly.