Using intelligent tutors based on Generative AI to strengthen programming skills in STEM and Non-STEM profiles

Abstract

The integration of Large Language Models (LLMs) into conversational tutoring systems can lower entry barriers in introductory programming courses with heterogeneous student profiles. This study evaluates the impact of such systems on performance and interaction, comparing students from STEM and non-STEM backgrounds. A quasi-experimental design ($N=30$) was implemented, comprising a control group (standard documentation) and an experimental group (intelligent tutor). Performance was assessed through five Python activities validated by unit tests, using a normalized score (0–1). The group utilizing TutorIA demonstrated a significant improvement, with average scores increasing from 0.43 to 0.78. Although STEM students maintained higher overall grades, non-STEM students achieved higher prompt quality scores (6.33/8.0 vs. 5.98/8.0), suggesting superior semantic contextualization. These findings indicate that TutorIA facilitated a leveling of the learning curve; however, both cohorts require distinct support mechanisms: logical reinforcement for non-STEM students and enhanced interaction strategies with AI assistants for STEM students.