Language Reshapes Experiential Learning in Humans and Machines

Humans acquire knowledge through two profoundly different routes: explicit instruction through language, and implicit experience accumulated through action. How these sources reconcile to guide decisions is poorly understood. To study their integration, we introduce the Instructed Bandit Tasks (IBT), eight sequential decision-making problems in which participants earn rewards while given natural-language hints conveying partial information about the task environments. Across two large (N=337 and N=300) behavioral experiments, linguistic hints improved performance and systematically reshaped choice trajectories. We propose modeling these effects with a hierarchical Bayesian RL agent, which treats language as a probabilistic constraint over beliefs, and compare it against Centaur, a language model fine-tuned on human behavior. Both models predicted which hints improve learning, and Centaur additionally reproduced when language hurt performance or lost its benefit under additional experience. Their complementary strengths suggest that normative and data-driven cognitive models can together capture the nuances of human learning.