Learning decouples accuracy and reaction time for rapid decisions in a transitive inference task

The accumulation of evidence over time, formalized in the drift diffusion model (DDM), has become a prevalent model of deliberative decision-making. Here, we adapted this model to understand better the role of latent variables in a serial learning task where decisions were made rapidly and did not show typical accuracy and response time patterns. We fit behavioral data using PyDDM (Shinn et al., 2020). We trained macaque monkeys (N = 3) on a transitive inference transfer task. They learned the implied order in a ranked list of 7 novel pictures in each behavioral session, indicating their choices by making saccadic eye movements. They reliably learned each new list order within 200-300 trials with asymptotic accuracy of around 80-90% correct. Their responses showed a symbolic distance effect, with 60% accuracy for adjacent list items and 90% accuracy for the largest symbolic distance. Although performance accuracy improved with learning and symbolic distance, reaction times were nearly invariant. Nevertheless, both accuracy and reaction time were well fit by the generalized drift-diffusion model. The fits were achieved by simultaneously increasing both the evidence accumulation rate and a collapsing bound to capture the shape of the reaction time distributions. These results indicate that decision-making during learning and transfer in a TI task may be characterized by a “variable collapsing bound” DDM. These results point to a unique dynamical regime of the DDM framework during serial learning.