Alexandra Neagu, Jeffrey T. H. Wong, Marcus Messer, Rhodri Nelson, Peter B. Johnson (2026). Pluralistic Alignment Workshop @ ICML 2026 📄 Full text (arXiv)
Key Findings
Real-world students show lower uptake of LLM tutor scaffolding than benchmarks assume (9,490 chats across 9 datasets). Bypassing scaffolding often signals a mismatch between pedagogical framing and learner goals, not a failure. Benchmarks must evaluate tutor adaptability to student-driven interaction.
Relevance to AI in Education
This paper contributes directly to understanding how AI systems interact with learners in authentic educational settings. Challenges benchmark assumptions about student uptake of LLM tutor scaffolding, showing real-world learners frequently bypass pedagogical framing for their own goals.
Related Pages
- scaffolding — related page
- intelligent-tutoring-systems — related page
- benchmark — related page
- llm-tutoring-feedback-diagnosis-gap — related page
- ai-fallibility-warning-help-seeking — related page
- correct-answer-trap-ai-tutor — related page
- didactical-teacher-assistant-dimensional-modeling — Encoding tutoring strategy in an explicit didactic layer (rather than implicit LLM prompting) makes
Citation
APA: Alexandra Neagu, Jeffrey T. H. Wong, Marcus Messer, Rhodri Nelson, Peter B. Johnson (2026). Rethinking Scaffolding in LLM Tutors: The Interactional Mismatch Between Benchmarks and Real-World Deployments. arXiv:2606.15766. Pluralistic Alignment Workshop @ ICML 2026.