Adaptive Behavior as Probabilistic Inference

Animals routinely face spatial decisions—where to forage, whom to follow, which refuge to choose—under uncertainty and time pressure, with limited energy and computation. Most theories of behavior address pieces of the problem—exploration, reward, or navigation—but few integrate them into a unified framework of information and action. I will introduce a theoretical framework that casts decision making as probabilistic inference under constraints: agents maintain beliefs over options and employ adaptive policies that couple movement and internal state to gather evidence and reduce uncertainty. I will begin by outlining how the same core mathematics unites adaptive systems across scales as a hierarchy of inference processes and formalizes a fundamental trade-off between exploration and exploitation—from the evolution of populations and societies to learning and decision making in individuals. Whether framed as maximizing fitness or reward, reducing uncertainty, or simply minimizing distance to a target, adaptive policies reduce to a small set of generic operations that drive individuals and populations along information gradients toward utility-driven goals. To illustrate these concepts, I will present a simple model of spatial decision making and show how this framework links top-down normative accounts of behavior with bottom-up emergent mechanisms, recovers familiar neural circuit motifs as special cases, and explains commonly observed behavioral patterns such as conflict, compromise, and commitment. Finally, I will connect this theory to data: using interpretable statistical models and inverse reinforcement learning, I will show how we operationalize the framework to address inverse problems by recovering policies directly from behavioral trajectories of animals moving through complex environments. I will discuss how these models enable counterfactual policies and causal feature attribution to identify which environmental cues reduce uncertainty in an animal’s choice of actions. Together, these results reframe adaptive behavior as the calculus of information—uniting inference and control in a theory whose consequences are encoded directly in the geometry of movement.