Computational Models of Dual-Process Reasoning

Dual-process theories posit that reasoning emerges from the interplay of fast, intuitive Type 1 processes and slower, deliberative Type 2 processes, regulated by metacognitive control. Yet despite their influence, these frameworks remain largely verbal and underspecified, limiting their capacity for falsifiable predictions. Here, we introduce a set of computational models that formalise key dual-process assumptions within an evidence-accumulation framework. Each model instantiates distinct mechanisms of working memory, inhibition, and confidence-based regulation and is evaluated against data from two-response bat-and-ball reasoning tasks. Six model variants—ranging from single-process “default” to confidence-regulated architectures—were compared using indices of accuracy, response time, and confidence. All models reproduced core reasoning signatures, but even the simplest captured many qualitative patterns, revealing that common verbal assumptions can be implemented by multiple mechanistic architectures. These findings underscore that while dual-process theories capture essential intuitions about intuitive–analytic dynamics, they remain underdetermined without formal specification. Computational modelling provides the necessary precision to advance dual-process theory from a descriptive dichotomy toward mechanistic explanation.