COG-RT · COGnitive latent dynamics of Reaction Times

Animals must adapt their behaviour in noisy environments under physiological and resource limits. Which latent cognitive processes drive these adaptations, and how they are implemented in cortical-striatal circuits central to control and motivation, remains incompletely understood. Classical cognitive models link choices and reaction times (RTs) to evidence, priors, or control demands, but they operate coarsely, averaging across conditions (e.g. costly vs. non-costly or model-based vs. model-free). This overlooks fine-grained, trial-by-trial strategy fluctuations that drive behaviour. Recent work shows that recurrent neural networks (RNNs) can recover hidden, continuous decision dynamics from behaviour that standard models miss. In the International Brain Laboratory (IBL) task, these advances reveal multi-timescale history effects and slow fluctuations in engagement without heavy parametric assumptions. These studies establish RNNs as powerful descriptive models of cognition, yet current models largely focus on choices, neglecting RTs - a complementary marker of engagement, urgency, and control allocation.This project will jointly model choices and RTs in the IBL dataset, containing hundreds of thousands of decisions with simultaneous whole-brain recordings, including anterior cingulate cortex (ACC) and striatum, enabling direct comparisons with neural activity. Integrating choices with RTs within cognitively constrained models will yield interpretable latent variables capturing dynamics of cognitive markers such as engagement, bias, and urgency, thereby describing decision content and timing in a unified way. Finally, we will test where and when these latents are encoded in ACC, striatum, and brain-wide activity. This project will provide a clear, testable picture of how hidden strategies shape decisions over time and where they live in the brain.