When a student hits a learning impasse, a tutor must balance two risks: intervening too early can suppress productive struggle, while waiting too long leaves the learner stuck in a frustrating loop. Modern generative AI tutors often employ guardrails that block direct answer provision, yet little is known about their behavior when an impasse persists. We examined 20,462 student turns from 1,260 authentic sessions with a guided LLM chemistry tutor, extracting 6,630 impasse turns classified into conceptual errors, expressed uncertainty, and help‑seeking. Using these impasses, we simulated three tutoring conditions—baseline, no‑direct‑answer, and guided tutor—and evaluated a sample of 150 impasses. Prompt specificity altered pedagogy: the baseline tutor gave the answer directly in 50.7% of cases, the no‑direct‑answer tutor asked a follow‑up question every time, and the guided tutor produced a wide range of context‑dependent responses. Analysis of real‑world impasse trajectories revealed that each additional impasse turn reduced the odds of recovery on the next turn by 12.7% (adjusted odds ratio = 0.873, p < 0.05).
Review