Large language models (LLMs) are increasingly employed in computer‑science education, yet even the most capable models suffer from pedagogical shortcomings: they tend to reveal solutions too early and lose focus on the learner’s immediate debugging goal. To tackle this, we introduce SocraticRepair, a planning‑guided Socratic tutoring framework aimed at program debugging and misconception repair. The framework is instantiated for Python by first synthesizing multi‑turn tutoring dialogues from four pre‑processed programming datasets, then fine‑tuning Qwen2.5‑Coder‑7B‑Instruct on these dialogues so that the model can generate a pedagogical plan, ask probing questions, and provide hints step by step, creating a Socratic interaction. Evaluation uses a set of predefined independent metrics—stepwise scaffolding, question quality, and premature solution disclosure—and benchmarks against a general‑purpose LLM. Results show that SocraticRepair achieves higher scores on scaffolding and question quality while markedly reducing early solution leakage. Nonetheless, generating reliable pedagogical plans remains a key open challenge, calling for further work to improve plan coherence and relevance.
Review