Answers are grounded in this paper and cite their sources.
After a failure, a generative tutor must decide how much help to give, yet current systems leave this to the language model.This paper introduces ReCal-GR, a framework that makes assistance intensity an explicit decision between learner-state prediction and generation.ReCal-GR estimates Recovery@3, the probability that a learner answers correctly within the next three opportunities sharing a concept and cognitive level with the failed item.A Recovery Calibration Index (RCI), formalized as a cost-sensitive Bayes rule over ordered assistance levels, maps this probability to an assistance region and a mandated action that constrain the generator.Its thresholds have a closed form, and any thresholds reveal the over-/under-assistance trade-off they imply.On 39,149 post-failure episodes from 6,463 MOOC-Radar learners, a calibrated logistic regression reached a ROC-AUC of 0.814 (95% CI 0.804-0.824)under learner-grouped cross-validation and robustly across horizons and cross-session windows.Because 88.6% of learners recover, tercile thresholds implicitly treat under-assistance as 9-23 times costlier than over-assistance and give the strongest support to 28% of learners who recover anyway.Cost-derived thresholds reduced this to 1.5%, lowered expected assistance cost by 0.24 (95% CI 0.23-0.25),and were never worse than terciles across 160 cost specifications.At an identical assistance-region distribution, a learner-history rule and a shuffled policy had higher cost (+0.06and +0.09).After a pilot exposed a saturated rubric, a blinded study of 360 Qwen2.5-3B-Instructoutputs (60 learners, three raters) found no difference in rated quality, but with the calibrated decision the realized intensity of help tracked predicted recovery (ρ = -0.57vs -0.02 without the decision and +0.22 with a shuffled one) and explicit answer disclosure fell from 11.7% to 0%.All code, derived data, cases, outputs and ratings are publicly released.
No takes yet. Share an insight, caveat, or question.
Khoual et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: