A Resource-Rational Mechanism for Reading

Reading requires the coordinated operation of perception, memory, and sequential decision-making under tight cognitive and temporal constraints. Yet a unified computational account explaining how these processes jointly shape eye movements and comprehension has remained elusive. Here we show that human reading behavior emerges from a resource-rational control mechanism in which readers allocate fixations to maximize expected gains in understanding while minimizing effort and time. We develop a hierarchical control model that formalizes reading as sequential information sampling under limited visual and memory resources. Implemented as a multi-level partially observable Markov decision process trained with deep reinforcement learning, the model links high-level comprehension with low-level eye-movement control within a single optimization framework. Across independent datasets and a new experiment involving time pressure, the model reproduces canonical behavioral patterns—including effects of word length, frequency, predictability, skipping, regressions, and coherence—while capturing adaptive changes in gaze and comprehension. These findings demonstrate that many canonical properties of reading arise from boundedly optimal information gathering and support a computationally unified account in which memory, perception, and control interact to guide intelligent behavior.