Task-space dimensions guide human exploration in complex environments

Abstract

Efficient exploration is essential for rapid learning in high-dimensional environments, where rewards depend on a sparse and unknown subset of task dimensions. Here, we developed a multidimensional value-learning task in which participants explored without explicit knowledge of the underlying structure. Across multiple task designs, participants consistently demonstrated successful learning under structural uncertainty. Behavioral analyses revealed that over half of participants adopted a structured exploration strategy--Full Dimension Scan (FDS)--with three key properties. First, participants systematically isolated and sampled one candidate dimension at a time, effectively reducing search complexity. Second, FDS was temporally persistent within dimensions while maintaining diverse sampling of feature values, enabling efficient local inference. Third, participants sequentially deployed FDS across dimensions, achieving broad coverage of the task space. Importantly, this strategy transferred across tasks: prior experience in lower-dimensional environments promoted stronger FDS use and improved performance in higher-dimensional settings. Consistent with these findings, a dimension-guided exploration model incorporating dimension-level attentional control best captured human behavior. Together, these results identify dimension-guided exploration as a core principle that transforms high-dimensional search into tractable, structured learning, providing new insight into adaptive behavior in complex environments.