NUMI

Computer-assisted learning in mathematics works best when students persist to mastery and learn from their mistakes. Most platforms let students stop early, and general-purpose AI tutors tend to hand over answers, which recent large-scale studies show can hurt learning.

NUMI delivers weekly, curriculum-aligned modules (a short video, exercises, an exit ticket, immediate feedback and worked solutions) with a mastery rule of three correct in a row and a guardrailed tutor beside every exercise. The tutor never gives the final answer. It offers a reasoning check to get started and, after a mistake, walks the student through the solution step by step, surfacing likely misconceptions. Student inputs are not stored by any external AI provider.

In a pre-registered experiment with 6,997 grade 6–8 students across 20 schools, the mastery requirement raised the share of students reaching mastery by 28.7 percentage points. The AI tutor raised the chance of answering correctly right after a mistake by 8.5 points on a 31.7% base. On a delayed test a week later, the tutor helped most when paired with a requirement to persist. Larger trials across grades 4–9 are under way.