Modeling hidden cognitive states reveals acute and chronic effects of fentanyl on decision-making.

The cognitive mechanisms underlying behavior are often dynamic, shifting gradually or abruptly over time scales spanning years, to weeks, to minutes. Whether drug-induced changes in learning and decision-making follow similarly dynamic patterns remains unclear. To address this, we apply a reinforcement learning model to choice data from rats performing a two-step task for oral fentanyl and sucrose rewards. The model contains a set of agents with their own learning and decision-making rules that differentially influence choice, and, critically, each agent's contribution to choice is allowed to vary across latent states that fluctuate over time. Using a dimensionality reduction method to align latent states across subjects, we identified three distinct states reflecting mixtures of goal-directed, habitual, and novelty-driven strategies. We found that acute fentanyl reward increased the frequency of transitions out of a goal-directed state into a habit-driven state, while chronic fentanyl exposure selectively diminished goal-directed control within a habit-dominant state, independent of sex. Together, these results demonstrate that fentanyl reshapes both the dynamics and cognitive architecture of decision-making, underscoring the utility of latent-state modeling combined with dimensionality reduction for uncovering drug-driven cognitive changes.