Widely used bandit tasks elicit diverging belief updating phenotypes in healthy adults

Adaptive decision-making requires dynamic arbitration between internal beliefs and incoming evidence, yet the role that environment structure plays in shaping these dynamics has received little systematic attention. We used three inductive reasoning and three bandit tasks to investigate these trial-by-trial dynamics in N=120 healthy adults. Using a novel computational model, we found that, in a two-option bandit task with positive valence, as well as in a three-option bandit task with negative valence, a significant number of participants were characterized by suboptimal arbitrations, reflecting increased reliance on priors under high uncertainty and increased reliance on evidence under high confidence. These findings demonstrate that design elements in widely used tasks can bias belief updating dynamics in healthy individuals leading to population-level suboptimal behaviors. These biases may have been overlooked in the past due to simplifications in the characterization of belief updating dynamics, potentially affecting analysis and comparisons with clinical cohorts.