Adaptive large language model scaffolding supports online programming education through learning analytics and knowledge tracing

Abstract

Online programming learners need timely support while managing heterogeneous prior knowledge and delayed feedback. This study evaluated a low-stakes adaptive design combining a knowledge-component map, event-history prediction, teacher-authored recommendations, a constrained large language model tutor, and explainable analytics. A quantitative-core mixed-method field evaluation compared adaptive and comparison conditions using identical core materials and assessments. The learning management system used stratified block allocation within four cohorts, but missing archival details required quasi-experimental interpretation. Data comprised 357 consenting learners, 8,568 pre/post item records, 6532 formative events, 1168 assistant sessions, and 42 follow-up responses. Pre-test and post-test reliability were KR-20 = 0.606 and 0.703. A regularised logistic event-history model achieved AUC = 0.683 (95% cluster-bootstrap CI [0.667, 0.698]); the logged knowledge-tracing probability achieved AUC = 0.660. The adjusted post-test difference was 3.07% points (95% CI [-0.87, 7.02], p = .127, partial eta-squared = 0.007). Adaptive learners showed a small raw-gain advantage of 4.26 points (95% CI [0.05, 8.47], Hedges g = 0.21), stronger formative and survey indicators, and unchanged completion. Teacher-governed large language model scaffolding was associated with stronger formative practice and perceived support, but not a significant improvement in the primary final academic outcome. It should remain a transparent, auditable, low-stakes aid rather than a grading or mastery authority.