Prior research has focused extensively on how emotion tendencies (e.g., duration, frequency, intensity, and valence) affect students' performance, but little is known about emotion variability (i.e., the fluctuations in emotion states) and how emotion variability affects performance. In this paper, emotion variability was examined among 21 medical students in the context of solving two patient cases of different complexity with BioWorld, a computer-based intelligent tutoring system. Specifically, we examined the influences of task complexity on emotion variability, emotion variability in self-regulated learning (SRL) phases, and the differences in emotion variability between high and low performers. We found that students' emotion variability varies depending on the SRL phases (i.e., forethought, performance, and self-reflection) and task complexity. High performing students had smaller emotion variability than low performers across the three SRL phases, but the differences in emotion variability were not statistically significant. Moreover, emotion variability in the forethought phase contributed most to high performance when compared to the emotional variability in the performance and self-reflection phases. This study advances theoretical development about emotion variability and provides insights that help explain the mixed results that existed in extant literature.