A controlled study of 66 graduate students in a mobile robotics course found that AI tutors using Socratic questioning methods produced superior learning outcomes compared to tutors focused on prompt refinement.

Researchers from EPFL tested two distinct AI tutoring approaches over six weeks. The Socratic-Guidance tutor structured interactions through dialogic questioning, while the Prompt-Refinement tutor taught students to formulate better prompts for language models.

Both tutoring methods achieved similar task performance during the initial guided phase. Students using either approach showed comparable prompting patterns and completion rates on programming assignments.

Learning transfer reveals key differences

The distinction emerged during a subsequent three-week project phase where 52 students used unconstrained LLMs without tutor guidance. Students who had worked with the Socratic tutor demonstrated higher learning gains and adopted more understanding-driven prompting strategies.

These understanding-focused prompts correlated with better comprehension of programming concepts. Students trained with Socratic methods were more likely to ask clarifying questions and seek explanations rather than direct code solutions.

Paradoxically, students perceived the Socratic tutor as less efficient during use. The prompt-refinement approach felt more streamlined and immediately productive to learners.

The study tracked actual prompting behavior rather than self-reported usage. Students with Socratic training showed measurably different interaction patterns when later using tools like OpenAI's models independently.

Implications for AI education tools

The findings suggest that immediate efficiency may conflict with long-term learning benefits in AI-assisted education. While prompt-engineering skills have obvious utility, the research indicates that dialogue-based scaffolding builds more durable learning capabilities.

The work appears in the proceedings of AIED 2026, where it received the conference's best paper award. The research team included educators from EPFL's computer science and robotics departments.

The study's methodology could inform the design of AI tutoring systems across programming education platforms, particularly as institutions grapple with integrating large language models into curricula.