Answers are grounded in this paper and cite their sources.
Long-horizon knowledge tracing requires more than an evocative memory mechanism: the implementation must assimilate observed responses non-anticipatorily, and the mechanism must survive matched controls. We subject Frac-BKT, a BKT-style model with an elapsed-time-scaled truncated fractional history, to both tests. The repaired model predicts response n from a prior mastery state and uses response n only to update the posterior state for later events. A five-seed evaluation uses 676,848 ASSISTments 2017 interactions from 1,694 learners and 19,249 learner--skill trajectories, with chronological 80/10/10 suffix splits. The response-aware model reaches AUC .5860 overall, .5638 after gaps of at least 30 minutes, and .5668 after at least 14 days. Relative to the earlier non-response implementation, the gains are .0473, .0316, and .0277 AUC, respectively. However, the fractional mechanism gate fails: the full model is not better than the matched $α=1$ control and is worse than the matched $τ=0$ control by .0090 AUC on both long-gap slices. Student-cluster wild-bootstrap intervals exclude zero, both Holm-adjusted p-values are .00006, and all five per-seed differences favor $τ=0$. Compact AKT-, SAKT-, and DAS3H-style references also remain stronger on long gaps. The contribution is therefore a response-assimilation repair and a falsifiable audit showing that response assimilation helps while the evaluated fractional-history term does not---not a state-of-the-art or causal-retention claim.
No takes yet. Share an insight, caveat, or question.
Nizam Kadir (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: