New Method Prevents LLM Tutors From Prematurely Revealing Answers

Large Language Model (LLM) based Socratic tutors, designed to guide students through questioning, often face a problem called "scaffolding collapse." This occurs when the tutor, under persistent student pressure, abandons its guided inquiry approach and directly provides answers, undermining the learning process. Existing solutions primarily focus on controlling the tutor's observable responses through prompting or filtering. However, these methods often fail to address the underlying internal representation drift that leads to this collapse. This paper proposes a novel framework, Scaffold-Preserving Representation Alignment, to tackle this issue. The framework involves a two-stage process: initial supervised fine-tuning, followed by a combination of trajectory-weighted direct preference optimization and a margin-preserving representation loss. This approach aims to maintain distinct internal states for scaffold-preserving versus collapse-inducing behaviors, thereby improving the robustness of long-horizon Socratic tutoring across various STEM disciplines and red-teaming attack strategies.