Learning to improve belief updating: a computational modelling approach - a follow up

Belief updating is often conceptualized as a static process in which individuals receive new information and decide whether to revise their beliefs. From this perspective, new evidence is assumed to exert a fixed influence on beliefs, largely independent of recent interaction history. In contrast, adopting an error-monitoring perspective, we propose that belief updating is a dynamic and learned process: individuals adapt how efficiently they integrate new information based on the structure of the informational environment and their prior interactions within it. A critical dimension of this environment concerns the type of information available. Some sources prioritize accuracy (e.g., official news agencies), whereas others emphasize agreement with existing beliefs (e.g., close peers who value social harmony). To formalize how individuals learn to navigate such environments, we implemented a reinforcement-learning–inspired framework in which the impact of new information depends on the interaction between its accuracy and its agreement with the individual’s current belief. Using a novel experimental paradigm, we independently manipulated the frequency of accurate feedback and belief-consistent (agreeing) feedback while holding objective task performance constant. This design allowed us to disentangle the informational value of accuracy from the motivational or social value of agreement, and to examine how participants dynamically adjusted their belief updating across trials and environments. Results from the initial study showed that accuracy feedback exerted a stronger influence on belief updating and belief change than agreement alone. However, this influence was not uniform. Instead, the effect of accuracy depended on recent interaction history: accuracy feedback was integrated differently depending on whether the peer had previously agreed or disagreed. Thus, accuracy information was weighted in a history-dependent manner rather than functioning as a static cue. At the same time, trial-level agreement robustly increased confidence, indicating that agreement strongly shapes confidence updating even when it does not directly drive belief revision. Despite these dynamic effects, we did not observe stable differences in belief updating between peers who predominantly agreed versus predominantly disagreed overall. That is, participants did not appear to form durable peer-type representations based solely on agreement frequency, even though agreement reliably influenced moment-to-moment confidence. This dissociation suggests that agreement may primarily affect confidence calibration rather than longer-term belief learning, or that agreement-based learning may be overridden by other task features. The present follow-up study is designed to clarify why stable differences between frequently agreeing and frequently disagreeing peers did not emerge. We consider two non-exclusive explanations. First, the explicit accuracy feedback provided at the end of each trial may have dominated learning, limiting the accumulation of agreement-based expectations about the peer. Second, agreement may boost confidence only up to a certain level, such that once confidence is sufficiently high, additional agreement no longer differentially affects belief updating, rendering overall agreement frequency less informative. By systematically examining the role of accuracy feedback and confidence levels in shaping belief updating, the follow-up study aims to identify the conditions under which agreement contributes to learning beyond transient confidence increases. More broadly, this work refines an error-monitoring account of belief updating by specifying when informational cues are integrated dynamically based on recent history rather than encoded as stable properties of social sources.