Reinforcement Learning algorithms allow agents to learn the probabilistic structure of the environment. These structures, termed belief-states, predict future outcomes and guide choice policies to maximize rewards. The ability to utilize new evidence to update belief-states is vital for adaptive behaviors, a process often impaired in schizophrenia contributing to psychosis, cognitive deficits, and avolition.A critical challenge persists in determining the underlying neuronal mechanisms governing belief-state updating and its impact on learning and reasoning. To address this, we train wildtype-mice in an adapted deterministic version of the human Two-step Markov Decision Task. Leveraging the significance of olfactory signals for rodents in the learning process, our current task utilizes odor stimuli to facilitate associations. In each trial, a fixed, but randomized pair of odors is presented. Odor presentation is directionally counter-balanced, and animals report choice by licking to the respective odor direction. LED lights indicate the deterministic transition to the chosen second state, followed by either a reward or no reward. As state transitions change randomly and unannounced in a session, animals need to adapt belief-states for optimal performance. Preliminary data indicate that mice effectively perform the task by discriminating between stimuli, selecting actions, evaluating outcome values, and improving their proficiency over time. We integrate sniffing and pupil dynamics tracking to assess task engagement and surprise, increasing the repertoire of potential cross-species markers. In the future, we aim to unravel neuronal mechanisms governing belief-state updating in both wildtype and schizophrenic risk mouse models.