In perceptual decision-making tasks, humans integrate noisy sensory evidence over time to guide their choices. The optimal integration process assumes that all evidence is weighted equally within a trial and that different trials are independent. However, humans exhibit systematic deviations from optimality, including uneven weighting of evidence within trials and influences from previous trials. Prior studies have demonstrated that biological constraints can account for this suboptimality. In this study, we present evidence that humans adapt their evidence integration strategies over time in response to task demands, and that the suboptimal uneven weighting is gradually learned over the course of the task. By explicitly modeling this adaptation through online gradient-based learning, our model outperforms existing approaches in capturing human behavior and unifies both observed forms of suboptimality in the Click task: dependence across trials emerges from an error-driven learning process that also gives rise to uneven integration weights within trials. We further propose a bounded-rational adaptation account to explain why humans progressively learn to weight evidence unevenly within a trial.

Our modeling framework provides a general approach of resource-rational adaptation. It captures how initially uninformed agents can gradually update their strategies through error-driven learning and is applicable to a broad range of learning and decision-making scenarios.