Tutoring software has a blind spot: it doesn't know what to do with a question it's never seen a student answer.
Researchers built a system called PICKT (Practical Integrated Cross-consistent Knowledge Tracing), detailed in an arXiv preprint (arXiv:2512.07179), to identify which data features keep a tutoring AI's proficiency estimates accurate when new questions carry no interaction history. The study combined several types of question features - including difficulty ratings, question text, and links drawn from a subject's knowledge map - and measured how much each contributed to accuracy on those unseen questions. Difficulty ratings mattered most for the hardest questions, the ones almost nobody answers correctly. Text and knowledge-map features helped differently, letting the model estimate a new question by finding older ones that were semantically or structurally similar.
Every intelligent tutoring system hits this problem the moment new content gets added: no student has answered the question yet, so the algorithm is guessing blind. The fix identified here isn't a fancier model - it's choosing the right metadata to feed it, a cheaper and more practical lever for the ed-tech companies actually running these systems day to day.
This is a narrower, more mundane cousin of the cold start problem recommendation engines have wrestled with for two decades, and the answer looks just as unglamorous: better labels, not bigger models.