<p>People use context to resolve ambiguous pronouns, for example, one contextual bias is to assign pronouns to the subject of the previous sentence, for example, Ana in "Ana ate with Liz. She…." Recent evidence suggests that these biases shift in the direction of the most frequent patterns in the local context. Yet referential adaptation is different from adaptation to words and structures, raising questions about the mechanism of adaptation. Here we test whether all recent exposures equally influence processing biases. Alternatively, learning may only occur when people encounter unexpected patterns, as predicted by error-based learning accounts that have been proposed for syntactic adaptation. We thus further test whether this “inverse frequency effect” also occurs during referential adaptation. In English the subject-reference pattern is more common than the nonsubject-reference pattern, so if people only adapt to rare patterns they should only adjust their processing biases when exposed to the less-common nonsubject-reference pattern. We tested this by comparing subject- and nonsubject-exposure with a neutral exposure condition. In two experiments we found that only the nonsubject-exposure condition differed from the neutral exposure condition, supporting the conclusion that referential adaptation exhibits an inverse frequency effect.</p>