In classical neuroscience experiments, neural activity is measured across many identical trials of animals performing simple tasks and is then analyzed, associating neural responses to pre-defined experimental parameters. This type of analysis is not suitable for patterns of behavior that unfold freely, such as play behavior. Here, we attempt an alternative approach for exploratory data analysis on a single-trial level, applicable in more complex and naturalistic behavioral settings in which no two trials are identical. We analyze neural population activity in the prefrontal cortex (PFC) of rats playing hide-and-seek and show that it is possible to discover what aspects of the task are reflected in the recorded activity with a limited number of simultaneously recorded cells (≤ 31). Using hidden Markov models, we cluster population activity in the PFC into a set of neural states, each associated with a pattern of neural activity. Despite high variability in behavior, relating the inferred states to the events of the hide-and-seek game reveals neural states that consistently appear at the same phases of the game. Furthermore, we show that by applying the segmentation inferred from neural data to the animals’ behavior, we can explore and discover novel correlations between neural activity and behavior. Finally, we replicate the results in a second dataset and show that population activity in the PFC displays distinct sets of states during playing hide-and-seek and observing others play the game. Overall, our results reveal robust, state-like representations in the rat PFC during unrestrained playful behavior and showcase the applicability of population analyses in naturalistic neuroscience. Display omitted •We used hidden Markov models to cluster neural activity of rats playing hide-and-seek•Inferred latent states follow the phases of the hide-and-seek game•Reverse physiology of latent states allows discovery of neurally relevant behaviors•Rat prefrontal cortex is in distinct sets of states while playing and observing play Bagi et al. use hidden Markov models to study PFC activity of rats playing hide-and-seek with humans. Such models, while naive and unsupervised, segment neural activity into meaningful states. States found in PFC activity alone correlate with hide-and-seek behaviors and allow for the discovery of previously undetected behaviors.