Prospective gains and losses influence cognitive processing, but it is unresolved how they modulate flexible learning in changing environments. The prospect of gains might enhance flexible learning through prioritized processing of reward-predicting stimuli, but it is unclear how far this learning benefit extends when task demands increase. Similarly, experiencing losses might facilitate learning when they trigger attentional reorienting away from loss-inducing stimuli, but losses may also impair learning by increasing motivational costs or when negative outcomes are overgeneralized. To clarify these divergent views, we tested how varying magnitudes of gains and losses affect the flexible learning of feature values in environments that varied attentional load by increasing the number of interfering object features. With this task design, we found that larger prospective gains improved learning efficacy and learning speed, but only when attentional load was low. In contrast, expecting losses impaired learning efficacy, and this impairment was larger at higher attentional load. These findings functionally dissociate the contributions of gains and losses on flexible learning, suggesting they operate via separate control mechanisms. One mechanism is triggered by experiencing loss and reduces the ability to reduce distractor interference, impairs assigning credit to specific loss-inducing features, and decreases efficient exploration during learning. The second mechanism is triggered by experiencing gains, which enhances prioritizing reward-predicting stimulus features as long as the interference of distracting features is limited. Taken together, these results support a rational theory of cognitive control during learning, suggesting that experiencing losses and experiencing distractor interference impose costs for learning.

Flexible learning of changing reward contingencies can be realized with different strategies. A fast learning strategy involves using working memory of recently rewarded objects to guide choices. A slower learning strategy uses prediction errors to gradually update value expectations to improve choices. How the fast and slow strategies work together in scenarios with real-world stimulus complexity is not well known. Here, we aim to disentangle their relative contributions in rhesus monkeys while they learned the relevance of object features at variable attentional load. We found that learning behavior across six monkeys is consistently best predicted with a model combining (i) fast working memory and (ii) slower reinforcement learning from differently weighted positive and negative prediction errors as well as (iii) selective suppression of nonchosen feature values and (iv) a meta-learning mechanism that enhances exploration rates based on a memory trace of recent errors. The optimal model parameter settings suggest that these mechanisms cooperate differently at low and high attentional loads. Whereas working memory was essential for efficient learning at lower attentional loads, enhanced weighting of negative prediction errors and meta-learning were essential for efficient learning at higher attentional loads. Together, these findings pinpoint a canonical set of learning mechanisms and suggest how they may cooperate when subjects flexibly adjust to environments with variable real-world attentional demands.

Attention includes processes that evaluate stimuli relevance, select the most relevant stimulus against less relevant stimuli, and bias choice behavior toward the selected information. It is not clear how these processes interact. Here, we captured these processes in a reinforcement learning framework applied to a feature-based attention task that required macaques to learn and update the value of stimulus features while ignoring nonrelevant sensory features, locations, and action plans. We found that value-based reinforcement learning mechanisms could account for feature-based attentional selection and choice behavior but required a value-independent stickiness selection process to explain selection errors while at asymptotic behavior. By comparing different reinforcement learning schemes, we found that trial-by-trial selections were best predicted by a model that only represents expected values for the task-relevant feature dimension, with nonrelevant stimulus features and action plans having only a marginal influence on covert selections. These findings show that attentional control subprocesses can be described by (1) the reinforcement learning of feature values within a restricted feature space that excludes irrelevant feature dimensions, (2) a stochastic selection process on feature-specific value representations, and (3) value-independent stickiness toward previous feature selections akin to perseveration in the motor domain. We speculate that these three mechanisms are implemented by distinct but interacting brain circuits and that the proposed formal account of feature-based stimulus selection will be important to understand how attentional subprocesses are implemented in primate brain networks.