Learning to Attend Through Value-Based Hypothesis Testing

In complex environments, humans must determine which features are relevant for learning and decision-making. Psychological theories offer competing accounts of this process: associative models suggest that attention emerges gradually through learned changes in feature values, whereas hypothesis-driven accounts propose that learners selectively attend to actively tested rules. Because attentional states are covert, similar behavior can arise from different underlying strategies, making these accounts difficult to distinguish using choice data alone. We inferred latent attention dynamics during learning and decision-making by training recurrent neural networks on synthetic data generated from feature-based reinforcement learning (FRL) and serial hypothesis testing (SHT) models. A network trained on hybrid (FRL+SHT) data outperformed single-model networks, decoding latent human attention with more than 80% accuracy. These results suggest that human attention reflects an interaction between value-based learning and hypothesis testing, in which learned feature value guides the generation and evaluation of candidate rules.