Exploration vs exploitation: how behavior shapes predictive representations in AI and mice

A new study published on arXiv (arXiv:2605.27929) by Kseniia Shilova and colleagues at Georgia Tech investigates how behavioral strategies—exploration versus exploitation—shape internal predictive representations in both artificial agents and biological mice. The researchers built an online learning agent navigating a tree-like maze using a predictive-coding framework. The agent continuously updated its perceptual model to predict future maze states and reward probability, with a controllable parameter that tuned the balance between selecting actions for information gain (exploration) or predicted reward (exploitation).

Results showed a striking divergence: exploratory agents developed latent representations that were highly spatially organized, preserving the structure of maze transitions. In contrast, exploitative agents learned more disorganized, reward-focused representations. To validate biological relevance, the team trained the predictive model on natural trajectories of water-deprived mice navigating the same maze. Mice that explored more exhibited representational geometries closely matching those of exploratory artificial agents, while mice with restricted visitation patterns aligned with exploiters. This work bridges computational neuroscience and AI, demonstrating that exploration fosters generalized, structured internal models—a principle that may guide the design of more robust autonomous agents and deepen our understanding of how animals learn spatial cognition.