Abstract
Understanding others’ beliefs and intentions is central to social interaction. Recent work has highlighted the role of individual characteristics (e.g., paranoia) in shaping how people interpret and predict others’ behaviour, yet most existing paradigms rely on simplified stimuli and assume shared or experimenter-defined world models. To address this limitation, we elicited participants’ beliefs about the distribution of two personality dimensions and four player-type clusters in Minecraft and then asked them to rate nine behavioural targets on these dimensions and estimate the likelihood that each target belonged to each cluster. After confirming that trait ratings predicted categorical judgments using multinomial logistic regression, we used participant-specific Bayesian mixture models that combined elicited beliefs with observed trait ratings to predict soft cluster assignments. We compared three model variants, assuming no correlation between traits, a general correlation, or cluster-specific correlations, and contrasted participant-specific (idiosyncratic) priors with group-level priors. Across multiple accuracy and similarity metrics, the idiosyncratic model without trait correlations consistently outperformed alternatives. These findings suggest that participants’ social categorisation is best explained by individualised belief structures operating over relatively independent trait dimensions, with apparent trait interactions emerging primarily at the group level rather than reflecting participant-level inference mechanisms.

This publication has 43 references indexed in Scilit: