In close relationships, we often maintain positive illusions about the relationship even in the face of negative information about the relationship partner. We propose that this phenomenon can be explained by a Bayesian inference model that groups partner behaviors by latent causes (such as their intentions and character). That is, the mental models can go beyond a faithful representation of a partner by integrating negative information into positive latent causes, in order to construct a positive illusion about the partner and the relationship. Using model simulations, we explore cases in which positive illusion construction is more likely to occur, focusing on the order of information (for example, a gradual versus sudden change in behaviors). This model provides a framework to reconcile how negative information about the partner can lead to bolstering the existing positive illusion in some cases and lead to change in impressions in other cases.