Inverse reinforcement learning reveals action-oriented value signals in naturalistic decision making

A major challenge for cognitive neuroscience is to explain how value of a goal-directed behavior is computed in complex and naturalistic environments. Standard computational models of decision making have been highly successful in controlled, trial-based paradigms, but they are often ill-suited to real-time behavior unfolding in naturalistic paradigms. Inverse reinforcement learning (IRL) offers a way to infer latent evaluative state from observed behavior in naturalistic environments, but its neural interpretability remains largely unknown. Here, we investigated whether moment-to-moment reward trajectories derived from IRL map onto value signals in the brain during a real-time driving task performed during fMRI scanning. IRL-derived reward trajectories were most robustly associated with activity in the dorsal striatum, a region often linked to value-guided action selection. They also showed associations with distributed regions supporting additional processes, including cognitive control and sensorimotor processing. This pattern suggests that IRL reward captures distributed neural activity centered on the reward circuitry, potentially reflecting how valuation interacts with other processes. Together, these findings suggest that IRL reward provides a behaviorally grounded, temporally resolved proxy for action-oriented valuation during naturalistic decision making.