Practice Less, Explain More: LLM-Supported Self-Explanation Improves Explanation Quality on Transfer Problems in Calculus

A research team from Carnegie Mellon University and the University of Washington has published a study demonstrating that Large Language Model (LLM) feedback can significantly enhance how students learn calculus. The paper, 'Practice Less, Explain More: LLM-Supported Self-Explanation Improves Explanation Quality on Transfer Problems in Calculus,' details a controlled experiment with 92 participants. Students were split into three groups: a control group with no self-explanation, a group using menu-based explanations, and a group using open-ended explanations with real-time feedback from an LLM. All groups showed learning gains, but the key finding was that quality, not just quantity, of practice matters.

While there was no significant difference in standard post-test scores, the open-ended LLM-feedback group excelled at a critical skill: explaining their reasoning on novel, complex problems. On challenging 'Not Enough Information' (NEI) transfer problems—where students must identify missing data—the LLM-supported group produced explanations that were 11.9 percentage points (or ~12%) higher in quality than the control group. This advantage emerged despite a major trade-off: students in this group solved substantially fewer practice problems within the same 60-minute timeframe. The study suggests AI can shift the focus from rote problem-solving to deeper conceptual understanding by providing scalable, personalized feedback on student reasoning, a resource typically limited by human tutor availability.

The findings, accepted at the AIED 2026 conference, indicate that integrating LLMs into educational tools could optimize learning efficiency. By automating high-quality feedback on explanations, AI allows students to spend more time on the metacognitive process of articulating their thought process, which proves more valuable for tackling unfamiliar problems than simply grinding through more exercises. This research provides a data-backed framework for developing next-generation 'learning companions' that prioritize depth over breadth.