When performing novel tasks, we often apply the rules we have learned from previous, similar tasks. Knowing when to generalize previous knowledge, however, is a complex challenge. In this study, we investigated the properties of learning generalization in a visual search task, focusing on the role of search difficulty. We used a spatial probability learning paradigm in which individuals learn to prioritize their search toward the locations where a target appears more often (i.e., high-probable location) than others (i.e., low-probable location) in a search display. In the first experiment, during a training phase, we intermixed the easy and difficult search trials within blocks, and each was respectively paired with a distinct high-probable location. Then, during a testing phase, we removed the probability manipulation and assessed any generalization of spatial biases to a novel, intermediate difficulty task. Results showed that, as training progressed, the easy search evoked a stronger spatial bias to its high-probable location than the difficult search. Moreover, there was greater generalization of the easy search learning than difficult search learning at test, revealed by a stronger bias toward the former’s high-probable location. Two additional experiments ruled out alternatives that learning during difficult search itself is weak and learning during easy search specifically weakens learning of the difficult search. Overall, the results demonstrate that easy search interferes with difficult search learning and generalizability when the two levels of search difficulty are intermixed.