A dual information seeking strategy improves imprecise human inferences outside the explore–exploit tradeoff

Everyday decisions require not just earning rewards but also learning about the world. We asked how people gather information when sampling is reward free, and compared this with reward seeking under identical outcome distributions. In a large study (N = 420), people employed two dissociable information seeking strategies. They often began by testing one option several times before switching, building early certainty, a sampling rule we call streaking. They also showed a global tendency to sample where uncertainty is greatest. Computational modeling shows that both strategies independently improve decision accuracy under noisy belief updating. Artificial neural networks trained to optimize performance acquired uncertainty-directed sampling but not early streaking, highlighting a feature of human sampling not spontaneously acquired by networks trained on these objectives. These results reveal a dual architecture of information seeking that links traits, sampling policies, and performance.