Inferring mind wandering from perceptual decision making.

People need to sustain focused attention to achieve goals. Yet, attention often lapses, as minds wander towards task-unrelated thoughts. The conventional way to study such shifts in attention is through thought probes that explicitly ask if thoughts are task-related. However, probes are rare and interrupt behavior. Other methods to measure mind wandering assume a 50/50 split in time spent on-task vs off-task. We address these issues with a framework to infer mind wandering (MW) using computational modeling. We use a random dot motion task with varying evidence, but with a strong bias inducing a repetitive response requirement. Occasional thought probes were used for validation. When participants (N = 93) reported being off-task, accuracy was higher and reaction time (RT) was lower, suggesting less stimulus processing and more reliance on bias. To classify internal states for individual trials from performance, we fit a Hidden Markov Model with Generalized Linear Models (GLM-HMM) for each state to responses. A two-state GLM-HMM predicted lower RTs on off-task trials, revealed an increase in mind wandering across the task, and aligned with self-reported focus. This shows that temporal variation in attentional states can be measured on a trial-to-trial basis without thought probes, paving the way for future MW research.