Learning Dynamics in Anxiety: The Role of Punishment Sensitivity and Learning Rate in Sequential Evaluation

Anxiety is fundamentally an anticipatory response to uncertain future threats, making sequential evaluation crucial. However, most computational research has involved one-step tasks that identify an elevated punishment learning rate as a key feature of anxious behavior, while some studies also suggest a lack of evidence for the role of the punishment sensitivity parameter. This contradicts recent research involving sequential evaluation tasks, which supports the idea that anxiety is primarily an uncertainty disorder. Nevertheless, these experiments have mainly focused on model-based algorithms and have left the role of the punishment learning rate unexplored. To reconcile the differences in the literature and better understand how these two parameters influence anxiety, two hybrid models were developed: one with differentiated learning rates for rewards and punishments, and another with differentiated sensitivity parameters. The models were then evaluated in the Cliff Walking task, using a deterministic environment and incorporating stochasticity in action selection to simulate an agent without complete control over its decisions. The results show that the impact of estimated punishments on planning is more significant than the speed at which they are learned, highlighting the central role of heightened punishment sensitivity in anxiety-related behaviors such as avoidance, risk aversion, threat overestimation and generalization.