In recent years, there has been a proliferation of adaptive learner models that seek to predict student correctness. Improvements on earlier models have shown that separate predictors for prior successes, failures, and recent performance further improve fit while remaining interpretable. However, students who engage in "gaming" or other off-task behaviors may reduce the predictiveness of learner models that treat counts of prior performance equivalently across gaming and non-gaming student populations. The present research evaluated how sub-groups of students that varied in their potential gaming behavior were differently fit by a logistic learner model, and whether any observed differences between sub-groups could inspire the creation of new predictors that might improve model fit. Student data extracted from a college-level online learning application were clustered according to speed and accuracy using Gaussian mixture modeling. Distinct clusters were found, with similar cluster patterns detected in three separate datasets. Subsequently, each cluster was separately fit to a Performance Factors Analysis model (PFA). Significantly different parameter coefficients across clusters implied that students more likely to have been gaming benefitted less from prior failures. These differences inspired new and modified predictors that were found to improve overall model fit - an improvement that varied in magnitude across clusters. The present findings indicate that incorporating trial duration into counts of prior failures can improve the predictive power of learning models. [This article was published in "Journal of Educational Data Mining" (EJ1230298).]