Patterns of belief perseverance and polarization reflect a resource-rational balance of epistemic and instrumental utilities

Belief polarization—the divergence of two individuals' beliefs despite encountering the same information—has been proposed to result from various mechanisms, including differing prior beliefs, trade-offs between accuracy and cognitive cost, and deviations from rationality due to motivated cognition. We propose a unified, resource-rational account of belief perseverance and polarization in which agents approximate Bayesian updating with a resource-limited particle filter model that adaptively resamples and rejuvenates a set of hypotheses upon observing evidence. By balancing epistemic objectives with the instrumental utility of holding a belief, the framework additionally captures the potential influences of motivated reasoning. We evaluate a family of resource-rational and motivated reasoning models on four past studies in which individuals exhibit belief polarization. Across studies, we find that models with modest particle counts and limited rejuvenation best reproduce human behaviour on the tasks; interestingly, this leads to polarization that is often more conservative at the group level than the predictions of normative models, despite the fact that a large proportion of individuals overreacts to the evidence relative to an ideal Bayesian observer. These results demonstrate that patterns of belief perseverance and polarization can be understood as rational compromises between epistemic accuracy and instrumental goals under bounded computational resources, offering a parsimonious mechanistic account of divergent belief updating across diverse domains.