Trial-by-trial variability in reaction time (RT) is ubiquitous in human behavior, yet its neural basis is poorly understood. In experimental studies RT variability is often modeled as random noise. While it may include a stochastic component, RT variability also reflects dynamic adjustments based on contextual factors such as whether or not the prior task was the same, performance history, and current contingencies 
. Here, we studied the neural basis of speed-accuracy trade-offs (SATOs) in response to errors. SATOs refer to trial-by-trial adjustments of RT that optimize performance. Such dynamic modulations of performance are fundamental to adaptive, flexible behavior. The SATO function depicts the non-linear relation between speed and accuracy such that faster responding does not affect accuracy, but only up to a point. Beyond that point speed and accuracy are inversely related, with slower responses having a greater probability of being correct (). That transition point can be regarded as an optimum, where the best accuracy is achieved at the fastest possible speed. SATOs are generally studied by manipulating participant instructions to emphasize either the speed or accuracy of responding. These manipulations are thought to lead to strategic changes in decisions about how much evidence is required to initiate a response. In the present study, rather than manipulating instructions, we examined SATOs in response to behavioral outcomes, specifically errors, which we hypothesize primarily reflect changes in task-directed attention. Over trials, responses speed up until an error is committed i.e., pre-error speeding; 
and following an error, RT slows [i.e., post-error slowing; 5], and the probability of an error decreases. This pattern can be interpreted as a progression to riskier positions on the SATO function culminating in an error, which is followed by a shift back to safer, more careful responding with a greater likelihood of a correct response.
Speed-accuracy trade-offs during the antisaccade task.
In the present study, we used functional MRI (fMRI) to test the hypothesis that SATOs in RT in response to errors are mediated by dynamic and reciprocal exchanges in the allocation of cognitive processing resources between the default and dorsal attention networks of the brain. Based on evidence that errors are characterized by transient declines in preparatory task-related attention 
and/or failure to suppress neural activity that interferes with performance 
and that attention is re-oriented in correct trials that follow errors 
, we expected activation in the default and dorsal attention networks to show reciprocal patterns of activation on trials surrounding an error, and to correlate in opposite directions with changes in RT. We used diffusion tensor imaging (DTI) to test the hypothesis that these network interactions rely on the integrity of the structural connections of the posterior cingulate cortex (PCC), which, as a cortical hub with a critical role in error processing, is in a privileged position to mediate network interactions around errors (evidence reviewed below). We expected greater coordination (i.e., reciprocity) between these networks and more pronounced SATOs in RT to correlate with greater microstructural integrity of the posterior cingulum bundle.
The default network is a set of functionally and structurally connected regions 
that typically show increased fMRI activation when not engaged in an overt task and deactivation during effortful task performance 
. The core regions of the human default network are the ventromedial prefrontal cortex, posterior cingulate/retrosplenial cortex, inferior parietal lobule, lateral temporal cortex, dorsal medial prefrontal cortex, and hippocampal formation 
. The default network is thought to be involved in self-referential and affective processing, and ‘task-induced deactivation’ is presumed to reflect the re-allocation of limited capacity cognitive processing resources away from the internal milieu to allow greater attention to the task-at-hand 
. Consistent with this theory, activation in the default network inversely correlates with activation in the dorsal attention network 
, which typically increases activity during cognitive tasks and is involved in effortful attention 
. During trials immediately preceding errors 
and error trials 
there is a relative failure of task-induced deactivation, suggesting that increasing interference from internally directed thought culminates in an error. Trials that follow errors, in contrast, show relatively increased task-induced deactivation and a corresponding increase in dorsal attention network activation 
, suggesting a shift in focus from the internal milieu back to the overt task. This cyclical pattern of network activation and deactivation around error trials is likely to be reflected in behavior. This is consistent with findings that the relative connectivity 
and activation 
of the default and dorsal attention networks correlate with RT. Building on this literature, we hypothesized that reciprocal changes in resource allocation between the default and dorsal attention networks around errors underlie SATOs in RT and are mediated by the PCC.
The PCC is the most densely connected cortical region, both structurally 
and functionally 
putting it in a privileged position to mediate network interactions. It is a major node of the default network and is negatively functionally connected to regions of the dorsal attention network. In the specific PCC subregions that show task-induced deactivation, functional connectivity with dorsal attention network regions is modulated by task difficulty, consistent with a role in regulating the balance between internally- and externally-directed attention that determines task-engagement 
. The PCC is also the putative generator of the error-related negativity (an event-related potential that indexes error detection) and the microstructural integrity of the posterior cingulum bundle predicts the speed of error self-correction 
. This suggests that the PCC detects errors and through its connections with other regions including the anterior cingulate and lateral prefrontal cortex 
, initiates immediate corrective action 
. Based on these findings that the PCC component of the default network is functionally connected with the dorsal attention network, that the strength of these connections is modulated by task difficulty, that the PCC detects errors and that the integrity of its white matter correlates with the speed of corrective action, we hypothesized that the PCC might also mediate network interactions around errors to support RT adjustments. Specifically, we hypothesized that the PCC mediates reciprocal changes in the balance of activity between the default and dorsal attention networks in response to errors. This would serve to reduce internal focus, increase attention to the task, and thereby support slower more careful responding to prevent error recurrence in subsequent trials.
To test our hypotheses that dynamic exchanges in resource allocation between default and dorsal attention networks underlie SATOs in response to errors and are mediated by the PCC we used an antisaccade task (), rapid-presentation event-related fMRI, and DTI. Antisaccades require inhibition of the prepotent response of looking towards a suddenly appearing stimulus and the substitution of a gaze in the opposite direction. Antisaccades are characterized by RT adjustments consistent with the SATO 
and errors (i.e., looking toward the stimulus) are associated with a failure of task-induced deactivation 
. We examined the correlations of task-related activation with RT over a series of five trials at different positions relative to an error: two trials pre-error (2PRE), one trial pre-error (1PRE), error (ERR), one trial post-error (1POST), two trials post-error (2POST). In the default network, we expected relatively increased activation on pre-error trials and a corresponding speeding of RT, followed by the fastest RT and greatest activation on the error trial, consistent with findings that antisaccade errors are faster than correct responses and are associated with greater activation in the default network 
. Following the error we expected a reinstatement of task-induced deactivation and a corresponding slowing of RT. In the dorsal attention network, we expected the opposite pattern (i.e., relatively decreased activation prior to and including the error followed by a post-error increase in activation) and positive correlations of activation with RT. This prediction is consistent with findings from both monkey single-unit recording and human neuroimaging studies of antisaccades suggesting that greater inhibition of the frontal eye field [reflected in reduced firing and increased fMRI signal, 27] is associated with lower error rates and slower correct responses 
. This pattern of network activation and RT changes would suggest that reciprocal changes in activation between default and dorsal attention networks underlie SATOs and reflect changes in the allocation of attention from the internal milieu to the task-at-hand. We also examined whether an index of coordination between these networks correlated with SATOs in RT. Finally, to test whether network coordination and SATOs rely on the structural connectivity of the PCC, we examined the relations of fractional anistropy (FA, a measure of white matter microstructrual integrity) of the posterior cingulum bundle with both the index of network coordination and post-error slowing. We expected greater integrity of the posterior cingulum bundle to be associated with greater coordination between the default and dorsal attention networks (i.e., more pronounced reciprocal pattern of activation) and more pronounced post-error slowing. In summary, we used fMRI and DTI to test the hypotheses that dynamic changes in attention to the internal versus external milieu in response to errors underlie SATOs in RT and are mediated by the PCC.