Decomposing response inhibition: a POMDP model

Abstract

Inhibitory control is a core cognitive function whose competence varies across the population, with impairments often observed in psychiatric conditions such as attention deficit hyperactivity disorder (ADHD). The Stop Signal Task (SST) is a widely used paradigm for assessing this ability. However, conventional formalizations of SST performance, such as the independent race model, rely on assumptions that are frequently violated in modern experimental designs. Furthermore, they typically fit only mean reaction times, overlooking crucial trial-by-trial dynamics. To address these limitations, we formalize the SST as a partially observable Markov decision process (POMDP). This framework characterizes inhibitory control with two components: noisy perceptual inference regarding stimuli and optimal control balanced against potential costs. To fit this model to the Adolescent Brain Cognitive Development (ABCD) study baseline cohort (N = 3,567), we introduce Transformer-encoded Simulation-Based Inference (TeSBI). This end-to-end architecture learns compact, sequence-aware embeddings from raw behavioral data. It enables efficient, amortized inference of individual-level posteriors. Extensive validation confirms it extracts reliable and identifiable parameters. We identify distinct latent computational attributes associated with scores on ADHD questionnaires. Controlling for sex, IQ, and medication status, children with higher ADHD scores exhibit subtle but robust shifts in computational attributes. They show a reduction in go cue directional precision, alongside a blunted sensitivity to stop error, go error and time costs. The learned embedding space reveals a continuous manifold in which children with higher ADHD scores are heterogeneously distributed, rather than forming distinct disorder clusters. This indicates that similar clinical characteristics can emerge from diverse combinations of computational mechanisms, supporting a dimensional perspective on neurodiversity. Our end-to-end framework can be extended to a broader range of cognitive tasks. It offers a scalable, theory-driven solution for analyzing large-scale behavioral data.