Listeners benefit from integrating prior beliefs with auditory sensory signals according to their relative reliability to reduce effects of stochastic noise. In volatile environments, however, listeners should not integrate when a change of signal source renders their prior beliefs irrelevant. Bayesian causal inference could provide the brain with an optimal strategy for adaptively adjusting its reliance on priors. Mechanistically, it has been suggested that the arousal system modulates the weights between prior and likelihoodaccordingly. We tested these hypotheses in an auditory localization task. Participants listened to noisy sound sequences of unknown length and with sudden changes in source location. Participants provided estimates of the last sound only, together with a spatial uncertainty interval, while we tracked their pupil size throughout stimulation as a proxy of arousal. We then fitted a Bayesian model to their behavioral responses and a deconvolution model to their ongoing pupil responses. Behavioral analysis indicated response biases towards predicted prior locations. In line with Bayesian inference, adaptive effects of prior reliability and relevance were clearly observed. Pupillometry analyses revealed negative correlations between modelled prior reliability and pupil baselines, and between prior relevance and evoked dilation. These results confirm that the brain has adopted principles of Bayesian inference to simultaneously manage stochasticity and volatility in perceptual decision making. In contrast to previous studies, the priors were formed without prompting to make explicit predictions, during ongoing and purely auditory stimulation. Moreover, our pupillometry results corroborate the modulating role of the arousal system during implicit perceptual belief updating.