From disengagement to re-engagement: behavioral evidence for AI-driven adaptive tutoring interventions in intelligent tutoring systems

Cognitive disengagement during intelligent tutoring system (ITS) interactions has been extensively characterized in terms of state detection and prevalence, yet the behavioral dynamics governing recovery from negative episodes remain poorly understood. This study analyzes cognitive-affective state dynamics in ASSISTments, a widely used mathematics ITS, using 597,242 interaction records from 799 secondary-level students collected over three academic years. A nine-phase analytical pipeline characterized state distribution, identified behavioral predictors of states through a three-stage feature selection procedure combining Kruskal–Wallis preselection, Random Forest importance ranking, and Recursive Feature Elimination with Cross-Validation, and modeled state transitions and recovery time from disengagement episodes toward the concentrating state. Results indicate that response time and hint-seeking behavior jointly account for 81.0% of Random Forest feature importance and exhibit large effect sizes in discriminating among states. Prevalence and severity were dissociated: boredom, the least frequent negative state, showed the lowest correct-response rate, while gaming, the most frequent, was also the most resistant to spontaneous recovery at the item level, though episodes that do exit resolve quickly given the brevity of individual gaming items. Recovery times differed substantially across negative states, ranging from a median of 18.0 s for gaming to 108.0 for boredom, roughly a six-fold difference. Average hint use during the negative episode emerged as the strongest behavioral predictor of fast recovery in the aggregate, with fast-recovering episodes showing 59% greater average hint use than slow-recovering episodes, although state-stratified analysis showed a positive association in frustration, a reversed association in confusion, a negligible association in gaming, and an initial positive association in boredom that reversed after filtering low-confidence assignments. These findings provide behavioral evidence for AI-driven adaptive interventions capable of promoting learner re-engagement and personalized support in Intelligent Tutoring Systems, shifting the focus from cognitive-state detection toward intervention strategies that sustain engagement.