Recent work using the Picture Word Interference (PWI) paradigm has revealed that language production, similar to non-verbal tasks, shows a robust Congruency Sequence Effect (CSE), defined as a decreased congruency effect following incongruent trials. Although CSE is considered an index of adaptive control, its mechanism is debated. In two experiments, we tested the predictions of a learning model of adaptive control in production, using a task-switching paradigm fully balanced to evaluate CSE on a PWI trial as a function of the congruency of a 2-back PWI trial (within-task CSE), as well as a 1-back trial belonging to a different task (cross-task CSE). The second task was a visuospatial task with congruent and incongruent trials in Experiment 1, and a self-paced reading task with ambiguous and unambiguous sentences in Experiment 2 that imposed a gap between the two PWI trials twice as long of that in Experiment 1. A learning model posits that CSE is the result of changes to the connection weights between task-specific representations and a control center, which leads to two predictions in our paradigm: (a) a robust within-task CSE unaffected by the intervening trial and the gap duration, and (b) an absent or reversed cross-task CSE. These predictions were contrasted with two versions of an activation model of CSE. In accord with the predictions of the learning model, we found robust within-task CSE in PWI in both Experiments with a comparable effect size. Similarly, evidence of within-task CSE was also found in the visuospatial and sentence reading tasks. On the other hand, examination of cross-task CSE from PWI to the other tasks and vice versa revealed either absent or reversed CSE. Collectively, these results support a learning model of adaptive control in language production.