AI Tutors Excel in Math Class with Human Oversight

In large university courses, students often face long waits for answers to their questions, especially during exam periods when instructors are overwhelmed. A new study demonstrates how artificial intelligence can provide immediate, accurate support while maintaining the essential role of human teachers. Researchers from the École polytechnique fédérale de Lausanne developed an AI system tailored to a Calculus I course, fine-tuning a lightweight language model on 2,588 historical student-instructor interactions. This approach addresses a critical need in education: scalable instructional assistance that doesn't compromise on quality or safety.

The system achieved a 75.3% accuracy rate on a benchmark of 150 representative questions, as evaluated by five experienced Calculus instructors. In 36% of cases, the AI's responses were rated as equal to or better than the instructor's original answers. The model performed strongest on medium-difficulty questions, which made up the majority of student inquiries, providing detailed explanations that instructors noted could be more helpful than brief human replies. However, for easy questions, it tended to over-explain, and for hard ones, it often failed to provide sufficient depth, revealing clear boundaries for its utility.

The researchers adopted a human-centered ology, working closely with the course instructor throughout development. They curated a dataset of 2,738 question-answer pairs from a course forum, translating them from French to English and augmenting them with intermediate reasoning steps to enrich training. The model, a fine-tuned version of DeepSeek R1 with 14 billion parameters, was selected after the instructor evaluated base models on 40 questions, prioritizing pedagogical alignment over raw performance. During deployment, the system included additional features like image-to-text conversion for handwritten questions and automatic classification of question types to tailor responses.

From a post-deployment survey of 105 students showed strong appreciation for the AI's immediacy and alignment with course materials. 69.5% reported that immediate answers saved them time, and 80% expressed interest in similar systems for other courses. However, trust was divided: 39% trusted the model enough to use its answers without always checking, while 31.5% disagreed. Instructor oversight proved crucial, with 44.7% of AI responses deleted due to pedagogical misalignment, though only 1.4% were edited, as instructors often preferred adding comments or rewriting entirely. The model's cost was approximately 56 times cheaper than leading commercial alternatives like GPT-5, making it accessible for institutional use.

The study's suggest a future where AI handles routine queries, freeing instructors to focus on complex or nuanced student needs. This hybrid model not only reduces workload but also helps students develop critical thinking by comparing AI outputs with human verification. Limitations include the focus on a single calculus course, which may not generalize to other disciplines, and the model's struggle with subjective aspects like relevance and completeness, where automated evaluation agreed with instructors only 55.3% of the time. The research underscores that AI in education should augment, not replace, human expertise, ensuring pedagogical rigor while embracing technological efficiency.