Category-like representation of statistical regularities allows for stable distractor suppression.

Statistical learning allows observers to suppress attentional capture by salient singleton distractors appearing at a predictable location. This form of learning is in many cases inflexible, persisting into extinction and across changes in probability without strong contextual cues. However, the underlying learning mechanisms and nature of distractor probability representations remain unclear. Here, we replicated learned location-based distractor suppression in two experiments, both showing a reduction in attentional capture when a salient distractor appeared at a high-probability distractor location, alongside impairments in target selection at that same location. We then used computational modeling to explore whether the strength of suppression was best explained by continuous distractor frequency summation, reinforcement learning prediction errors, or a categorical all-or-nothing response. In both experiments, our data were most parsimoniously explained by a combination of a global exponential decay in response time with each distractor presentation paired with a categorical learning mechanism in which the most highly associated location was suppressed over all other locations. Our results suggest that the magnitude of suppression is more sensitive to overall probability differences than to subtle trial histories.