Individual differences reveal shared computational structure in perceptual decision-making

Abstract

To discover computational mechanisms of cognition, models typically specify a fixed psychological architecture, treating individual differences as variations within a predefined parameterization. Here, we take a different approach by using the population structure of individual differences to directly drive model discovery. We develop a low-rank neural network that uses this `population geometry' to learn shared computational components and participant-specific weights from trial-level choices, allowing cross-subject variation to constrain the model's architecture. Applied to perceptual decision-making in the Bernoulli Clicks Task across primary and replication datasets, out-of-sample prediction reveals a six-dimensional representation for individual differences: two continuous dimensions of temporal evidence weighting and two of choice-outcome history dependence, alongside participant-specific side bias and signal-to-noise ratio. This low-dimensional organization replicated in an independent dataset. Together, these findings reveal that seemingly disparate decision heuristics reflect different weightings over a small set of shared computational components, and establish a principled methodology for using behavioral variability to discover the mechanistic architecture of cognition.