Unravelling Repetitive Negative Thinking With Reinforcement Learning

Recent advances in the computational dynamics of planning and state inference from the interdisciplinary field of reinforcement learning offer rich opportunities for insights into repetitive negative thinking (RNT), specifically rumination and worry. In this perspective, we apply key principles from reinforcement learning and meta-reasoning to provide a normative foundation for clinical phenomena associated with RNT, including excessive focus on potential negative events, the impact of overly abstract thinking, and the perpetuation of RNT over time. We explore how these factors may contribute to clinically relevant behavioral outcomes such as avoidance. We propose two algorithmic accounts of RNT: worry-as-planning and rumination-as-inference, where agents learn through mentally simulating states and actions. Furthermore, we discuss how these algorithms can be viewed as cognitive actions subject to selection, learning, and reinforcement. This integration opens avenues for innovative approaches to understanding and intervening in maladaptive thought patterns, ultimately advancing the treatment of RNT-related conditions.