A common way of quantifying learning behavior is the estimation of learning rates from subjects’ responses. To this end responses are usually averaged across subjects when estimating learning rates for a whole group or averaged over a sliding window when estimating learning rates for single subjects. However, these methods are difficult to apply when the number of experimental subjects and/or the number of trials are too low. Furthermore, the process of averaging obscures the actual dynamics of learning, which can occur within few trials. Indeed, we found that for previous data from a contextual reversal learning task (Uengoer and Lachnit, 2006) that during learning single subjects do not gradually change responses, but appear to abruptly switch their responses. In this task, subjects were shown stimuli, i.e., food items, in either of two contexts, i.e., restaurants, in each trial and had to predict whether the presented item will result in stomach troubles. We analyzed the responses for all stimuli by identifying behavioral change points (CPs) using binary segmentation. Change points during acquisition were biased towards the first two trials and did not differ significantly between stimuli. During reversal CPs shifted towards later trials and were more variable indicating that acquisition memory interfered with reversal. Consistent with these results we found that for data from a follow-up study (Uengoer, Klass et al., 2020) that the backward shift of CPs during reversal is larger when reversal occurs in the same context. These data are not easily accounted for with simple associative learning models such as the Rescorla-Wagner model. Hence, based on previous work (Batsikadze et al., 2022), we used a deep Q-network (DQN) to model the choice behavior of subjects in the reversal learning tasks reported by Uengoer & Lachnit (2006) and Uengoer, Klass et al. (2020), and compared it to a classical Rescorla-Wagner model. Both models were trained on the experiences of each subject and a simple grid search was used to determine the hyper-parameters, e.g., learning rate, which yielded the best fit. While both models were able to learn the tasks and exhibited a similar CP distribution only the DQN was able to learn context-dependent responses. In summary, we show the importance of analyzing subject-level learning dynamics, show the utility of change point analysis and test the ability of different models to account for learning behavior.