Large language models are increasingly used in computer science education, but their teaching instincts are flawed. According to a new arXiv paper, even the most powerful LLMs tend to reveal the solution prematurely and may lose focus on the learner's actual state. That undermines the point of tutoring, where guiding a student to find their own error is more valuable than handing over a fix.

The paper, titled "SocraticRepair: Planning-Guided Socratic Tutoring for Program Debugging and Misconception Repair," proposes a planning-guided framework that keeps the tutor on track. Instead of letting the model free-associate an answer, SocraticRepair structures the dialogue around a plan for debugging, using Socratic questioning to lead the learner toward identifying and repairing their own misconceptions.

The work targets a practical gap: debugging is a core skill in programming education, but LLM-based tutors often fail at it pedagogically. By explicitly planning the tutoring session, the authors aim to make LLM tutors more reliable partners in the classroom. The abstract does not include evaluation results, so the effectiveness of the approach remains to be seen, but the problem it addresses is well defined.