Weighting Bayesian information in human judgment under uncertainty: a hierarchical theory of cognitive biases.

Cognitive biases are typically viewed as departures from normative Bayesian reasoning. Instead, we propose that such biases emerge from a hierarchical cognitive architecture that adaptively regulates the weighting of prior beliefs and incoming evidence when cognitive resources are limited and uncertainty exists. While existing approaches, including heuristic, ecological rationality, Bayesian cognition, and resource-rational theories, capture important aspects of probabilistic inference, they differ in how inferential weighting is specified and adapted across contexts. We introduce Adaptive Bayesian Cognition (ABC), a hierarchical computational framework in which informational channels based on prior beliefs and evidence compete dynamically during belief updating. ABC distinguishes three interacting strata: first-order inferential updating in log-odds space; second-order reliability learning, which is driven by prediction error; and third-order adaptation to environmental uncertainty. Within this architecture, inferential weighting emerges from the competitive arbitration of informational sources based on attentional salience, cognitive cost, learned reliability, and inferred environmental uncertainty. ABC reconceptualizes cognitive biases as context-sensitive consequences of adaptive inferential arbitration, rather than as failures of rationality. Importantly, this framework makes a unique dynamical prediction: salience effects on evidence weighting should vary based on environmental uncertainty and learning history. This yields selective crossover dynamics that are not predicted by other accounts. Overall, ABC provides a unified computational framework that links probabilistic inference, reliability learning, and uncertainty adaptation in resource-constrained systems across ecological environments. Thus, ABC offers a computational framework of cognitive biases as systematic consequences of context-dependent evaluations of prior information and evidence in uncertain environments.