Students’ ability to self-evaluate is crucial for learning and academic success. However, global explicit self-evaluations (self-reports; e.g., self-assigned grades) are prone to biases. Here, we examined whether eye movements in learning quizzes can be used as a trial-based, implicit measure of self-evaluation and predict exam success. Participants’ answers to learning quiz questions triggered a lateralized positive/negative feedback after a short delay (anticipatory interval between answer and feedback). Crucially, per participant, correct feedback was always displayed on one side of the screen and error feedback on the opposite side. During the anticipatory interval between answer and feedback, question and answer options either remained on screen or disappeared (display condition). This manipulation allowed us to investigate whether the absence of visual stimulation was essential to infer implicit self-evaluations from anticipatory saccades, eye movements to the correct/error feedback side prior to feedback appearance. We used a signal detection approach to assess the (mis)match between participants’ anticipatory saccades and the correctness of their respective quiz answers. Corresponding signal detection measures were then used to predict learning quiz performance and later exam success using linear mixed models. Sufficient numbers of anticipatory saccades occurred for both display conditions. Yet, signal detection measures derived from anticipatory saccades failed to predict exam success in both display conditions, whereas global explicit self-evaluations (self-assigned grades) significantly predicted exam success. We discuss why trial-based and implicit measures of self-evaluations seem to fail at predicting future exam success.