Abstract: Simulated learner systems are used for many purposes rang- ing from computational models of learning to teachable agents. To sup- port these varying applications, some simulated learner systems have relied heavily on machine learning to achieve the necessary generality. However, these efforts have resulted in simulated learners that sometimes make generalization errors that the humans they model never make. In this paper, we discuss an approach to reducing these kinds of generaliza- tion errors by having the simulated learner system reflect before acting. During these reflections, the system uses background knowledge to recog- nize implausible actions as incorrect without having to receive external feedback. The result of this metacognitive approach is a system that avoids implausible errors and requires less instruction. We discuss this approach in the context of SimStudent, a computational model of human learning that acquires a production rule model from demonstrations.