Evaluating large language models in Calculus learning: Performance, prompting strategies, and conceptual understanding
Year: 2026
Type: article
Study design: Clinical Trial
Abstract: This study investigates the pedagogical effectiveness of Large Language Models (LLMs) in calculus education through a mixed-methods quasi-experimental design comprising three sequential stages and four interconnected sub-studies examining mathematical problem solving, prompting strategies, and conceptual understanding. Although artificial intelligence (AI)-powered conversational systems are increasingly integrated into STEM education, robust empirical evidence regarding their instructional effectiveness in higher mathematics remains limited. To address this gap, the study benchmarked ChatGPT-4o, Gemini 2.0, and LLaMA 4 using 90 calculus problems across multiple conceptual domains, examined the effects of Zero-Shot, Few-Shot, and Chain-of-Thought (CoT) prompting on model accuracy and explanatory clarity, and conducted a comparative classroom experiment involving 75 undergraduate students to evaluate conceptual understanding of the epsilon–delta definition of limits. The findings demonstrate that prompting strategy is a critical determinant of LLM performance, with CoT prompting consistently producing the greatest reasoning coherence and explanatory clarity, although performance varied across calculus topics. Integral problems were solved most consistently, whereas limit problems posed the greatest challenge. Statistical analyses further showed that increasing mathematical complexity frequently reduced explanation clarity despite correct numerical solutions. Students receiving ChatGPT-assisted instruction demonstrated significantly different conceptual understanding outcomes than those receiving conventional lecture-based instruction (χ² = 6.726, df = 2, p = 0.035, Cramér’s V = 0.299, 95% CI [0.111, 0.505]), with fewer students remaining at minimal comprehension and more achieving partial understanding. Moreover, the logistic mixed-effects model indicated that Few-Shot prompting significantly reduced explanation clarity relative to CoT prompting (β = −0.222, 95% CI [−0.312, −0.132]). Interpreted through Cognitive Load Theory, these findings suggest that CoT prompting reduces extraneous cognitive load by making intermediate reasoning explicit. Overall, this study provides empirical evidence supporting the responsible integration of generative AI into undergraduate calculus instruction by emphasizing prompt literacy and reaffirming the indispensable role of human guidance in fostering conceptual understanding, proof-oriented reasoning, and higher-order mathematical thinking.
Source: Journal on Mathematics Education
Language: English
Subfield: Education
Field: Social Sciences
Domain: Social Sciences
SDG: Quality education
Open Access status: gold
Estimated APC paid: $250
OpenAlex ID: W7171293575
