We introduce a new model of adaptive criterion setting withina signal detection framework, and show how this provides psy-chological insights that allow us to segregate causes of subop-timality in perceptual learning. We apply this to a perceptuallearning task for both neurotypical and autistic participants.The model parameters provide a bridge between the mecha-nisms of an aberrant precision account of autism and result-ing behavior that can be interpreted within a receiver operatingcharacteristic framework. The model makes superior out-of-sample predictions compared to standard signal detection the-ory, about how people adapt to different environmental manip-ulations when asked to categorize audio-spatial stimuli. Wefind that accuracy of participants is more strongly correlated tothe construct of persistence signals that inhibit response flexi-bility, than to the neuromodulatory gain. We also find evidencefor individual differences in persistence that are correlated toscores on the autistic traits questionnaire.