A fundamental property of natural intelligence is the capacity to recognise and leverage statistical patterns in the environment to promote adaptive behaviour. This capability, often termed statistical learning (SL), enables us to construct accurate models of the world that aid in perception and decision-making, when faced with noisy and ambiguous sensory input. However, despite extensive data from human experiments and proposals from theoretical neuroscience, the neural implementation of SL remains elusive. To address this, we employed a two-alternative forced-choice (2-AFC) sound categorisation task in head-fixed mice. Mice classified white noise volumes as louder or quieter than an implicit midpoint boundary. After achieving expert performance on a uniform distribution, we expose them to different sensory priors by manipulating sound distributions such that one of the categories was oversampled near the boundary, while keeping the likelihood of the two categories fixed at 50\%. A normative model, capturing perceptual noise, predicts a systematic shift in the subjective estimation of the boundary away from the locally oversampled area, such that the performance around the boundary becomes biased toward the oversampled category, causing a horizontal shift in the psychometric curve. Moreover, such optimal agent would show greater shift with higher sensory noise. Both of these predictions were borne in the data. We then implemented a generalised linear model-hidden Markov model (GLM-HMM) with a novel unsupervised method to determine the optimal number of latent behavioural states. This allowed us to exclude states where animals exhibited idiosyncratic, stimulus-independent strategies and to identify behavioural modes corresponding to different sensory priors. Building on these findings, we investigated the neural mechanisms underlying exploitation of different sensory priors using optogenetic inactivation. Our results suggest that the anterior cingulate cortex (ACC) plays a crucial role in utilising sensory context priors during decision-making processes.