Mitigating Scaffolding Collapse in Socratic Tutors via Representation Alignment
Prior defenses primarily constrain observable responses through prompting, preference optimization, or filtering, leaving the internal representation drift that precedes trajectory-level collapse largely unaddressed. We propose Scaffold-Preserving Representation Alignment, a two-stage framework that first warms up a Socratic tutor with supervised fine-tuning, then combines trajectory-weighted direct preference optimization with a margin-preserving representation loss anchored to frozen reference states. Our method is designed to maintain separation between scaffold-preserving and collapse-inducing hidden states across dialogue turns.
- ▪Prior defenses primarily constrain observable responses through prompting, preference optimization, or filtering, leaving the internal representation drift that precedes trajectory-level collapse largely unaddressed.
- ▪We propose Scaffold-Preserving Representation Alignment, a two-stage framework that first warms up a Socratic tutor with supervised fine-tuning, then combines trajectory-weighted direct preference optimization with a margin-preserving repre
- ▪Our method is designed to maintain separation between scaffold-preserving and collapse-inducing hidden states across dialogue turns.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2607.19371 (cs) [Submitted on 15 Jun 2026] Title:Mitigating Scaffolding Collapse in Socratic Tutors via Representation Alignment Authors:Jing Shao, Qifeng Wu, Hanyu Zhang, Sixia Sun, Jun Zhuang View a PDF of the paper titled Mitigating Scaffolding Collapse in Socratic Tutors via Representation Alignment, by Jing Shao and 4 other authors View PDF HTML (experimental) Abstract:Large language model (LLM)-based Socratic tutors increasingly guide students through multi-turn questioning, but they can suffer from scaffolding collapse: under sustained student pressure, a tutor gradually abandons guided inquiry and reveals solutions directly.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
