TutorMoments: Do AI tutors know when to help and when to hold back?

TutorMoments is a framework designed to evaluate if large language models can effectively balance helping a student versus letting them work through a problem. The system uses real, teacher-annotated math tutoring transcripts to identify key decision points where an LLM takes over as the tutor in a simulated session. An evaluation pipeline then scores the model's performance on whether it appropriately provides support (scaffolding), pushes the student for deeper thinking (rigor), or over-helps. Preliminary results show that models tend to over-help by default but improve when prompts explicitly define the pedagogical trade-off, though they still differ widely and have room for improvement.