Learning to improve belief updating: a computational modelling approach

Belief updating is typically viewed as a static process, where individuals receive new information and make decisions on whether to update their beliefs. Adopting an error monitoring perspective, we propose that individuals can learn to integrate new information more efficiently, leading to more accurate beliefs. Integration efficiency depends on information type. Some environments provide accurate information (e.g., official news agencies), while others prioritize agreement over accuracy (e.g., a close friend valuing your feelings). Reinforcement learning (RL) models align well with our error monitoring approach to belief updating, demonstrating how agents optimize rewards instead of strictly following unbiased updating rules. Implementing a RL approach, we assume that the impact of new information on a current belief depends on the interaction between its accuracy and its agreement. A new research design is developed that systematically varies the frequency of accurate and agreement information with the current belief while keeping objective performance constant. The basic paradigm can easily be extended to understand learning with motivated beliefs and the involvement of confidence levels.