Humans learn from mismatches between expectations and observations. Yet, what counts as a prediction error depends on what is mentally represented – which features are considered relevant and how the environment is believed to be structured. We demonstrate how mental model formation constrains subsequent learning and gives rise to a trade-off between adaptability and effort-minimisation.

In two experiments using a hierarchical multi-armed bandit task, participants were assigned to one of two conditions in which either an initially irrelevant feature became predictive (flexible condition) or an initially predictive feature lost relevance (constrained condition) midway through the task. We hypothesised that participants in the flexible condition would filter out the initially irrelevant feature from their representation and thus be blind to an emerging regularity; while participants in the constrained condition would retain the irrelevant feature in the second half and explore its new role.

We show that different representational policies – operating under the same error-driven mechanism – produce distinct outcomes consistent with the proposed trade-off: filtering of the initially irrelevant feature increased early efficiency but masked subsequent change in this feature, whereas entertaining a richer representation (i.e., believing the irrelevant feature is relevant in some yet-to-be-discovered way) enabled adaptive insight but required greater exploratory effort. Surprisingly, insight prevalence was comparable across conditions, suggesting that the tendency to simplify or entertain richer models does not solely depend on environmental statistics but individual differences. Our results highlight how mental model structure shapes adaptation.