We need to be flexible to adapt to dynamically changing circumstances. A probabilistic reversal learning task is one ofthe experimental paradigms to characterize flexibility of a subject. In recent studies, it is hypothesized that a subject mayutilize not only a reward history but also a cognitive map representing a latent structure of the task. In this study, weconducted an experiment using the task toward understanding a process of learning a latent structure of the task. We foundsubjects choose a rewarding option with relatively high frequency in a later phase of the task. Analyzing the subjectsdecision making, it is suggested that they make decision based on their own estimation about the latent structure. Astatistical model selection suggested that a reinforcement learning model with state representations fit behavioral data inthe later phase. These results suggest the subjects learn the latent structure during the task.