QuickTA

One-to-one tutoring produces some of the largest learning gains known, and it has never been affordable at scale. In a course of a thousand students, help arrives late, unevenly, and mostly in office hours. Generic chatbots fill the gap by handing over answers.

QuickTA is a course-aware conversational tutor built on large language models with pedagogical guardrails: it withholds complete solutions, stays on the assignment’s topic, and follows a hint-then-check structure drawn from worked-example and metacognitive-questioning research. After assignments it runs LLM-guided reflection to support self-regulated learning, and every interaction is logged for research. A campus-hosted version runs on small language models with retrieval-augmented generation, so student questions never leave university hardware.

More than 15,000 post-secondary students have used QuickTA in different capacities, in courses spanning computer science, mathematics, statistics, chemistry, engineering, and writing. Across deployments in introductory databases and CS1 courses at the University of Toronto, 54% of students’ first queries were conceptual rather than solution-seeking. Brief guidance on how to use the tutor cut verbatim copy-pasting, trust in the tutor rose over the term, and guided reflection raised self-reported confidence, although exam-score differences were not statistically significant. In a cluster-randomized study in a 774-student mathematical statistics course, giving the tutor access to course materials did not change tutorial scores, but international students and students with lower final grades were more likely to find it useful.