Default mode network activation at task switches reflects mental task-set structure

The default mode network (DMN) is a set of regions in the human brain that show elevated activity at rest, contrasting with attenuation during tasks that demand externally-focused attention (Buckner et al., 2008). Prominent views suggest that the DMN may be engaged during construction of internal, especially autobiographical, mental models, or during exploratory monitoring of the environment in the absence of attentional focus. This pattern of responding typically contrasts with networks such as the multiple-demand network (Duncan, 2010) which are active in attentionally demanding situations, and often anti-correlate with the DMN (Fox et al., 2005). Subsequent research showed that the set of default mode regions can be further separated into “Core,” “MTL,” and “dMPFC” subnetworks (Andrews-Hanna et al., 2010, 2014; Andrews-Hanna, 2012; Axelrod et al., 2017; Wen, Mitchell, et al., 2020), which have partially specialised roles in various internally-focused cognitive processes, with the Core subnetwork particularly implicated in self-referential processing (Andrews-Hanna et al., 2010; Gusnard et al., 2001; Kelley et al., 2002), the MTL subnetwork involved in autobiographical memory and imagery (Addis et al., 2007; Palombo et al., 2018; Schacter et al., 2007), and the dMPFC subnetwork linked to social cognition (Denny et al., 2012; Gallagher et al., 2000; Hassabis et al., 2014; Saxe & Kanwisher, 2003; Tavares et al., 2008).

Complicating this broadly consistent picture, DMN activity can sometimes be observed during goal-directed and stimulus-driven situations (Spreng, 2012), especially when responses are relatively automatic (Vatansever et al., 2017). Here, we are particularly interested in recent findings of increased DMN activity, especially of its Core subnetwork, during difficult, externally cued task switches, which further challenge simple views of DMN as a purely “task-negative”, or internally focused network (Crittenden et al., 2015; Kurtin et al., 2023; Smith et al., 2018; A. X. Zhou et al., 2024). These studies are based on the classic behavioural paradigm in which cued changes in the required response mapping are associated with behavioural performance costs (Allport et al., 1994; Egner, 2023; Kiesel et al., 2010; Monsell, 2003; Vandierendonck et al., 2010), but have employed a larger-than-typical set of tasks. Studies requiring people to alternate between two, often similar, tasks typically activate multiple-demand-like regions, rather than the DMN (Braver et al., 2003; Kim et al., 2012; Monsell, 2003). The studies reporting DMN activity during task-switching have instead used sets of at least four tasks, often with a hierarchical organisation. Two of these studies (Crittenden et al., 2015; Smith et al., 2018) used six tasks, grouped into three stimulus domains, and found that Core and MTL subnetworks of the DMN responded to between-domain switches, but not to within-domain switches. More recently, we investigated DMN activity as people switched between four tasks, grouped into two domains (A. X. Zhou et al., 2024). Although we replicated a DMN task-switch effect, showing increased Core DMN activity at task switches versus task repeats, we found comparable DMN activity for all types of task transitions, including within-domain switches. The reduced number of domains in the latest study suggests that the number of tasks may influence whether the DMN is recruited by task switches, and further raises the question of the role of the DMN in externally-driven cognitive transitions.

To investigate whether the DMN task-switch effect depends on the complexity or organisation of the set of tasks, we therefore designed a paradigm that preserves the task-switching trial structure, while manipulating the set of tasks in two critical ways. Firstly, since the number of tasks may constitute a threshold for the involvement of DMN, we compared runs in which participants had to switch between eight different tasks (from four stimulus domains) with runs in which participants were only asked to perform four of these tasks (from two stimulus domains). Secondly, to test whether the DMN might also be sensitive to higher-order, hierarchical mental organisation of the set of tasks, we had participants learn and practise the tasks in two groups of four, allowing us to examine switches between learnt groups and effects of learning order.

We hypothesised, firstly, that if the Core DMN task-switch response depends on the number of tasks to be performed, then within-domain and between-domain switches would differ more in runs with four domains (replicating Crittenden et al., 2015; Smith et al., 2018) than in runs with two domains (replicating A. X. Zhou et al., 2024). Secondly, we hypothesised that if the Core DMN is sensitive to higher-order task structure, then this could be reflected in a greater univariate response for between-group switches compared to within-group switches, a different response to tasks learned later compared to tasks learned first, and a multivariate response pattern that distinguishes tasks from different learning groups better than those from the same learning group.

Thirty-eight native English speakers between the age of 18–45 (64.7% female), and with normal or corrected-to-normal vision and no history of neurological or psychiatric disorders, were recruited from the Cognition and Brain Sciences Unit’s healthy participant panel. Two participants were excluded from analysis due to technical errors in presenting stimuli during the experiment and saving data. Participants all provided informed consent and were given monetary compensation for their time. The experiment was conducted in accordance with ethical approval granted by the Cambridge Psychology Research Ethics Committee (CPREC).

Estimated effect sizes were not available based on existing literature, so the choice of sample size assumed a medium standardised effect size of Cohen’s d = 0.5 for the key hypothesis that Core DMN’s within-domain and between-domain switch responses would differ more in four-domain than two-domain runs. A two-tailed t-test with alpha of 0.05 and power of 0.8 yielded a minimum required sample size of 34. Secondary hypotheses are orthogonal, and the same logic and sample size apply. Thirty-eight participants were recruited on the expectation that some may have needed to be excluded.

The experimental paradigm, adapted from Crittenden et al. (2015) and Smith et al. (2018) and depicted in Figure 1, comprised eight different colour-task pairs, with the pairings fixed across participants. Each task was cued by a different coloured frame. Within a sequence of trials, cues could repeat, or change to a different colour that indicated a different task, as in many traditional “task switch” paradigms (Kiesel et al., 2010; Monsell, 2003; Vandierendonck et al., 2010), but with a large and hierarchically structured set of tasks. The colours red and green framed an object, and cued “Is it living?” and “Does it fit into a shoebox?” respectively. The colours blue and yellow framed an incomplete word and respectively cued the lexical tasks “Does I fit in to make a word?” and “Does A fit in to make a word?” The pink and cyan colours framed a face and indicated “Is this person female?” and “Is this person young?” Finally, the orange and purple colours framed two shapes and asked “Are these the same shape?” and “Are these the same height?”

On each trial, a coloured frame (4.36 × 4.36 degrees of visual angle) appeared in the middle of the screen, surrounding a simultaneously presented stimulus and cueing the task to be performed. Rest trials were indicated by a black frame, with the participant not required to respond. Each trial ended when the participant responded, or, for rest trials, after the mean reaction time of all preceding trials. The next trial followed after an inter-trial interval of 1.5 s, with a central fixation cross.

Participants learnt these eight different task types, two in each stimulus domain, in two practice sessions before entering the scanner. Each session included learning and practising the colour-cue and task rules for two of the four domains. Eleven participants learnt the word and face tasks first, followed by the object and shapes tasks; 12 participants learnt the object and shapes task first, followed by the word and face tasks; and 13 participants learnt the object and word tasks first, followed by the faces and shapes tasks. During this pre-scan training, participants were first shown the colours that corresponded to each rule, listened to the experimenter describe the rules, and read the instruction sheet until they were confident that they had learnt the pairings. Then, participants practised the tasks cued by the coloured frames, one colour/rule pair at a time, for ten repeats. Participants were then asked to repeat out loud which question each colour cued and to complete a block of 30 trials with mixed cues and a random order of trials. Participants were allowed to proceed to the testing phase if their performance on the practice block exceeded 80%, otherwise they attempted another practice block. After learning the first four tasks in this way, the participant moved on to a similar practice session for the second set of four.

In the main experiment, across each run, we pseudorandomised the sequence in which the tasks were presented, to create different switch conditions (Fig. 1B). For two-domain runs, switch types were task repeat, within-group-within-domain switch, within-group-between-domain switch, task-to-rest switch, and rest-to-task switch (“restart”). In four-domain runs, there were additionally between-group-between-domain switches. Within each run, equal numbers of trials per switch condition were counterbalanced across task types. Participants completed four runs with tasks from only two domains, in the same pairs they were learnt, and two runs with trials intermixed from all four domains (Fig. 1C). Each two-domain run had 81 trials, 16 for each of five types of task switch (see above), plus a dummy trial with a random task at the start of the run; each four-domain run had 193 trials, with 32 for each of six types of task switch, following an initial dummy trial.

In the first three runs, each task was restricted to be followed by only one of the possible tasks per switch type. For example, if the task on trial N was “Is it living?” (Fig. 1A), then for trial N+1, a task repeat trial would also be “Is it living?”, a within-group-within-domain switch trial would be “Does it fit into a shoebox?”, a within-group-between-domain switch trial would always be the same one of the two possibilities, for example, “Does A fit in?”, and in a four domain run, a between-group-between-domain switch trial would again always be only one of the four possibilities (e.g. “Is person female?”). In this way, the probability of encountering a particular task transition was fixed across trial types. In the second three runs of the experiment, this constraint was removed, so that, for a given trial type, any of the possible next tasks could occur. fMRI findings were very similar for the two halves of the session, so all results from the two were averaged.

Data were acquired on a 3T Siemens Prisma MRI scanner fitted with a 32-channel coil. A structural T1-weighted structural scan was acquired using an MPRAGE sequence (TR 2.4s, TE 2.2 ms, flip angle 8°, voxel size 0.8 × 0.8 × 0.8 mm). Task runs used T2*-weighted Echo-Planar Imaging (multiband acquisition factor 2, TR 1.2 s, TE 30 ms, flip angle 67°, voxel size 3 × 3 × 3 mm).

Pre-processing involved spatial realignment, slice-time correction, co-registration, and normalisation to the MNI template brain (using SPM12 and AutomaticAnalysis; Cusack et al., 2015). For both univariate region of interest (ROI) and multivariate analyses, no spatial smoothing was used.

Within the DMN, we focused on the Core subnetwork because this has shown the largest and most consistent response to task switches in previous studies (Crittenden et al., 2015; A. X. Zhou et al., 2024). The Core DMN ROI was taken from Wen, Mitchell, et al. (2020), based on local voxel clusters from the Yeo et al. (2011) 17-network parcellation that contain the Core subnetwork’s coordinates provided in Andrews-Hanna et al. (2010). (See Wen, Mitchell, et al., 2020 for further details.)

For each participant, a general linear model (GLM) was created, with regressors for each combination of switch condition (task repeat, within-group-within-domain switch, within-group-between-domain switch, between-group-between-domain switch, restart, rest), number of task domains in the run (two or four), task expectancy (constrained transitions, all transitions), and learning group (first or second). Note that some combinations did not exist in the experiment or model: between-group-between-domain switch could not occur in two-domain runs, and rest trials did not belong to either learning group. Regressors were created by convolving the response periods¹ (from stimulus onset to trial end, indicated by the participant’s response, or for a matched duration on rest trials), of all trials per condition, with the canonical hemodynamic response function. Six rigid-body realignment parameters were included as non-convolved regressors per run, along with run means. The model and data were temporally high-pass filtered using the default cut-off period of 128 s. Analysis focused on the mean signal across the Core DMN ROI, and compared switch conditions using repeated-measures ANOVAs and paired t-tests. ANOVAs also tested whether effects of switch type interacted with two potential modulators: current task set complexity, operationalised as the number of domains included in the run, and instructional complexity, operationalised as whether each task had been learnt in the first or second group of domains (Fig. 1D). Note that some comparisons involved trial types that were present in all runs (e.g., the various within-learning-group switch conditions) whereas others compared trial types across runs (e.g., regressors from four-domain runs versus regressors from two-domain runs). The regressor for between-group-between-domain switches only existed in four-domain runs, so it was compared to within-group-between-domain switches from four-domain runs but not from two-domain runs. Frequentist statistics were supplemented with default Bayes factors based on F statistics for ANOVAs (Faulkenberry & Brennan, 2023) and t statistics for t-tests (Rouder et al., 2009).

Neural discrimination of each pair of tasks was examined by training linear support vector machine (SVM) classifiers (LIBLINEAR, Fan et al., 2008) to classify tasks, for each ROI, and each participant, using The Decoding Toolbox (Hebart et al., 2015). Response patterns were beta coefficients from a second GLM that was similar to the first except that it had separate regressors to model all trials of each task type per run. Cross-validation was performed across runs, testing separately on held-out four-domain and two-domain runs, and training on all possible combinations of remaining runs for which classes and domain number were balanced. That is, each classifier was trained with data from two runs, one two-domain run and one four-domain run, and tested on the held out runs. Classification performance was calculated separately for each test run type (two-domain or four-domain), and for each relationship between task pairs: either coming from the same domain, from different domains within the same learning group, or from different learning groups. In each case, classification performance was averaged across classifications conducted using all possible combinations of training runs, and was quantified as the mean signed distance of test patterns from the decision boundary (with incorrect classifications having a negative sign). This measure was chosen because it gives a continuous, unbounded, and unbiased estimate of representational distance, while being insensitive to differences in pattern variance between conditions (Nili et al., 2020). Classifier performance was compared to chance using two-tailed t-tests, and compared across switch conditions using repeated-measures ANOVA and paired two-tailed t-tests.

Mean reaction time (RT) for correct trials (Fig. 2A) was highly similar in two-domain runs (1.64 s), and in four-domain runs (1.75 s), with no significant difference (t₃₅ = 0.56, p = 0.58, BF₁₀ = 0.21). Switch costs were evident, where average reaction times for task switches (combined across types; correct trials only) were significantly higher than task-repeat trials in both the two-domain runs (t₃₅ = 9.85, p < 0.01, BF₁₀ = 7.33 × 10⁸) and the four-domain runs (t₃₅ = 9.92, p < 0.01, BF₁₀ = 8.72 × 10⁸).

Reaction times of correct trials were examined further using a three-way, repeated-measures ANOVA with factors of switch type (task-repeat, within-group-within-domain, within-group-between-domain, and restart), number of domains present in the current run (two or four), and the learning group to which the task belongs (first or second). Switch type showed a significant main effect (F_(3,105) = 91.13, p < 0.01, BF₁₀ = 8.71 × 10²⁵), while other main effects and interactions were non-significant (F < 1.33, p > 0.26, BF₁₀ < 0.39).

Post hoc, paired t-tests across switch types found significantly higher reaction times for within-group-within-domain switch trials compared to task repeat trials (t₃₅ = 7.75, p < 0.01, BF₁₀ = 2.93 × 10⁶), and significantly higher reaction times for within-group-between-domain switch trials than within-group-within-domain switch trials (t₃₅ = 11.40, p < 0.01, BF₁₀ = 3.10 × 10¹⁰). While restart trials had significantly higher reaction times compared to within-group-within-domain switch trials (t₃₅ = 8.23, p < 0.01, BF₁₀ = 1.16 × 10⁷), no significant difference was observed between restart and within-group-between-domain switch trials (t₃₅ = 1.97, p = 0.06, BF₁₀ = 1.00).

To investigate effects of task switches between tasks from different learning groups, a two-way, repeated-measures ANOVA using only trials from the four-domain runs assessed factors of switch type (within-group-between-domain, between-group-between-domain) and learning group. There was no significant effect of switch type, learning group, or interaction (F_(1,35) < 0.70, p > 0.41, BF₁₀ < 0.30).

Mean accuracy was consistently high, at 96% across all trials (Fig. 2B). A three-way ANOVA with factors switch type, domain number, and learning group, similar to that conducted for RT, confirmed no significant differences between conditions (F_(3,105) = 2.58, p = 0.07, BF₁₀ = 0.13), domain numbers (F_(1,35) = 0.01, p = 0.90, BF₁₀ = 0.21), or learning groups (F_(1,35) = 1.09, p = 0.30, BF₁₀ = 0.35). A two-way ANOVA with only trials from four-domain runs and with factors switch type (within-group-between-domain and between-group-between-domain) and learning group showed no significant main effects or interaction (F_(1,35) < 0.06, p > 0.80, BF₁₀ < 0.22).

In the Core DMN ROI, we first examined the basic switch effect, contrasting against task-repeat trials, and averaging across all other factors (Fig. 3A). Replicating previous findings of the DMN switch effect (Crittenden et al., 2015; Kurtin et al., 2023; Smith et al., 2018; A. X. Zhou et al., 2024), a t-test confirmed a significant increase in Core DMN activity at task switches (averaged across switch types) compared to task repeats (t₃₅ = 4.46, p < 0.01, BF₁₀ = 301). A t-test also found greater Core DMN activity for within-group-between-domain switches compared to within-group-within-domain switch trials (t₃₅ = 4.43, p < 0.01, BF₁₀ = 280).

Next, we tested whether responses to between-group-between-domain transitions were greater than for within-group-between-domain transitions in four-domain runs, as evidence for a higher-order hierarchy in the switch effect. However, matching the behavioural results, transitions between different learning groups did not significantly differ from transitions between domains within a group (t₃₅ = 0.30, p = 0.76, BF₁₀ = 0.19).

Also in line with behaviour, the response at transitions from rest back to task (restart) was significantly greater than at within-group-within-domain switches (t₃₅ = 2.73, p = 0.01, BF₁₀ = 4.24), but not significantly different from within-group-between-domain switches (t₃₅ = 0.85, p = 0.40, BF₁₀ = 0.25) or between-group-between-domain switches (t₃₅ = 0.57, p = 0.57, BF₁₀ = 0.21).

Finally, we compared the average of all task-switch trials to rest trials. This confirmed that the Core DMN ROI was significantly deactivated by the task trials compared to the rest trials, as expected for a DMN region (t₃₅ = 9.73, p < 0.01, BF₁₀ = 5.38 × 10⁸).

We next examined the effect of each hypothesised modulator of the task-switch effect, using repeated-measures ANOVAs, in which transition types of interest (within-group-within-domain and within-group-between-domain) were crossed with the two modulating factors of current task-set complexity (two domains, four domains) and instructed order (learnt first, learnt second). Since some studies (Crittenden et al., 2015; Smith et al., 2018) had found a difference between within-domain and between-domain switches, but our most recent study had not (A. X. Zhou et al., 2024), we first focused on these two transition types, and were specifically interested here in their interaction with each potential modulating factor. To ensure that effects of learning order did not depend on differential responses to particular stimulus domains, these ANOVAs also included a between-participant factor of which two domains were learnt first (words and faces, objects and shapes, or words and objects).

Contrary to prediction, there was neither a significant main effect of domain number (F_(1,33) = 0.09, p = 0.77, BF₁₀ = 0.23), nor an interaction of domain number with transition type (F_(1,33) = 1.14, p = 0.29, BF₁₀ = 0.37), suggesting that increasing the complexity of the current set of tasks did not increase the difference between switch types (Fig. 3B). Interestingly, however, there was an interaction of switch type with the learned order of the domains (Fig. 3C), with higher activity for between-domain switches of later-learned domains (F_(1,33) = 5.23, p < 0.05, BF₁₀ = 2.03). There was no main effect of learned order (F_(1,33) = 0.42, p = 0.52, BF₁₀ = 0.27). Paired t-tests revealed no significant difference between within-domain and between-domain switch conditions within the first learnt group (t₃₅ = 1.45, p = 0.16, BF₁₀ = 0.46), replicating A. X. Zhou et al. (2024), but a significant increase in Core DMN activity in between-domain compared to within-domain conditions for the second-learnt group (t₃₅ = 3.92, p < 0.01, BF₁₀ = 72.72), replicating Crittenden et al. (2015) and Smith et al. (2018). The between-participant factor of domain-pairing had no significant main effect (F_(2,33) = 1.54, p = 0.23, BF₁₀ = 0.24), nor interactions with factors of transition types, domain number, or learned order (F_(2,33) < 0.91, p > 0.41, BF₁₀ < 0.14), indicating that the effect of learning order generalised across the particular group of tasks learnt first.

Next, to confirm that the Core DMN response to between-domain switches is similar whether within or between groups, even when accounting for learning order, we ran a second ANOVA using only between-domain switches from four-domain runs and crossing transition type (within group, between group) with instructed order (learnt first, learnt second). The types of domains learnt first (words and faces, objects and shapes, words and objects) were again added as a between-subject factor. There was a main effect of instructed order (F_(1,33) = 5.80, p = 0.02, BF₁₀ = 2.53), with activity higher for later-learned tasks. Neither the main effect of transition type nor interaction was significant (F_(1,33) < 3.87, p > 0.06, BF₁₀ < 1.18). Again, there was no significant effect of the between-subject factor of type of domains learnt first (F_(2,33) = 0.86, p = 0.43, BF₁₀ = 0.13), nor any interaction with transition type or instructed order (F_(2,33) < 1.47, p > 0.24, BF₁₀ < 0.22).

To assess whether the distinctions between tasks, domains, and learning groups were represented within the Core DMN, we performed multivariate pattern analysis to distinguish every pair of tasks. In Figure 4, the average task-pair discrimination strength is plotted, separately for task pairs that come from the same domain, from different domains within the same learning group, or from different learning groups, and separately for two-domain and four-domain runs. A three-by-two ANOVA with factors of task-pair relationship and domain number found significant main effects of domain number (F_(1,35) = 9.03, p < 0.01, BF₁₀ = 8.53), task-pair relationship (F_(2,70) = 194.33, p < 0.01, BF₁₀ = 6.48 × 10²⁵), and a significant interaction (F_(2,70) = 11.10, p < 0.01, BF₁₀ = 294).

Post hoc two-tailed t-tests confirmed stronger discrimination of within-group-between-domain task pairs than within-group-within-domain task pairs, for two domain runs (t₃₅ = 17.36, p < 0.01, BF₁₀ = 5.37 × 10¹⁵), and also for four domain runs (t₃₅ = 15.53, p < 0.01, BF₁₀ = 1.91 × 10¹⁴), consistent with previous results (Crittenden et al., 2015; Smith et al., 2018) and as expected based on perceptual differences between stimulus domains. Extending this result, two-tailed t-tests revealed stronger discrimination of between-group-between-domain task pairs than within-group-between-domain pairs, both for two-domain runs (t₃₅ = 2.73, p < 0.01, BF₁₀ = 4.28) and also for four-domain runs (t₃₅ = 8.29, p < 0.01, BF₁₀ = 1.25 × 10⁷). This implies that the Core DMN also represents the abstract task structure imposed by the learning groups. Interestingly, a two tailed t-test showed that between-group task pairs were also discriminated better in four-domain runs (when tasks from both groups were performed), than when tested on trials in the two-domain runs (t₃₅ = 3.33, p < 0.01, BF₁₀ = 16.66).

Because domains were assigned to learning groups such that the face tasks were never learnt in the same group as the object or word tasks, it was important to confirm that greater discrimination of between-group task pairs could not be due to greater dissimilarity of these particular task combinations. Therefore, the comparison of between-group-between-domain pairs versus within-group-between-domain pairs was repeated, but excluding those task pairs that were always between-group, and adding the between-participant factor of first-learnt domains (words and faces, objects and shapes, words and objects). The within-participant factor of domain number was also included. This three-way mixed ANOVA showed no significance for the between-participant factor or any interaction with it (F <1.15, p > 0.33, BF₁₀ < 0.17). The significant main effects of both domain number (F_(1,33) = 7.05, p = 0.01, BF₁₀ = 4.08), and task-pair type (F_(2,66) = 191.49, p < 0.01, BF₁₀ = 5.22 × 10²⁴), were retained, along with their interaction (F_(2,66) = 8.82, p < 0.01, BF₁₀ = 53.62). Therefore, the greater neural discrimination of tasks learnt in different groups generalised across the particular identity of these tasks.

Finally, we assessed whether discrimination of within-group-within-domain or within-group-between-domain task pairs depended on whether the tasks were from the first or second learning groups. In the first learning group, participants were assigned either the object and word domains, object and shape domains, or word and face domains; thus this analysis was restricted to the latter two subsets, for whom domains were balanced across first and second learning groups. A three-way mixed ANOVA included within-participant factors of task-pair type (within-group-within-domain, within-group-between-domain) and learning group (first, second), plus the between-participant factor of first-learnt domains. Aside from the previously reported effect of task-pair type (F_(1,21) = 110.59, p < 0.01, BF₁₀ = 4.23 × 10⁶), there was no significant effect of learning order (F_(1,21) = 6.46 × 10^-5, p = 0.99, BF₁₀ = 0.28), and no significant interaction of task-pair types and learning order (F_(1,21) = 2.12, p = 0.16, BF₁₀ = 0.67). There was also no significant main effect of the between-participant factor of types of domains learnt first (F_(1,21) = 4.26 × 10^-3, p = 0.95, BF₁₀ = 0.28) and no significant interaction with task-pair types (F_(1,21) = 0.35, p = 0.56, BF₁₀ = 0.33), although a significant interaction with learning order (F_(1,21) = 21.08, p,<0.01, BF₁₀ = 147) suggested that task discrimination depended on the identity of the tasks.

There is transdiagnostic interest in the functional connectivity of the DMN, with abnormal DMN coupling reported for various clinical conditions (Assaf et al., 2010; Y. Zhou et al., 2007). Correspondingly, it is possible that the functional connectivity of the Core DMN differs as a function of switch condition. Although we did not have a priori hypotheses in this regard, we conducted exploratory analyses (see Supplementary Materials) following the suggestion of a reviewer. We used beta-series regression (Rissman et al., 2004), based on trial-wise vectors of activation estimates per condition, to examine two types of connectivity: intra-network connectivity between the anterior and posterior nodes of the Core DMN, and connectivity between the whole Core DMN and the multiple demand network (MDN), a set of frontoparietal regions that are typically activated by attentionally demanding tasks (Duncan, 2010; Fedorenko et al., 2013), and that are expected to have an antagonistic relationship with DMN regions during such tasks (Fox et al., 2005).

As expected, anterior and posterior nodes of the Core DMN were positively correlated, while Core DMN and MD were negatively correlated (Supplementary Figs. 1 and 2). Correlations were closely similar across conditions, except that the anterior-posterior correlation within the DMN was somewhat increased on rest and restart trials (Supplementary Figs. 1 and 2) and decreased for four-domain runs.

This study set out to investigate factors that affect recruitment of default mode regions at external task switches, and to resolve an inconsistency across studies, whereby equal Core DMN response to within-domain and between-domain task switches has been observed with four tasks organised into two domains (A. X. Zhou et al., 2024), but preferential response to between-domain switches has been observed with six tasks organised into three domains (Crittenden et al., 2015; Smith et al., 2018). Here, we report imaging results that replicate higher activation of Core DMN at task switches overall, along with a tendency for higher activity on between-domain shifts modulated by overall complexity. Specifically, we found no evidence that the switch effect differed for four-domain runs compared to two-domain runs. Instead, we found a significant interaction between learning group and task switch type. For the tasks learnt first, Core DMN activity was similar for within- and between-domain switches, replicating A. X. Zhou et al. (2024). For the tasks learnt second, activity was greater for between-domain switches, replicating Crittenden et al. (2015) and Smith et al. (2018). These results suggest that the Core DMN is sensitive to a mental task structure established during task learning. With accumulating complexity of the total set of tasks, the later tasks become increasingly chunked by domain, such that Core DMN becomes preferentially sensitive to coarser, between-domain task transitions.

The insensitivity to the complexity of the current set of tasks is also reflected in the behavioural results. Reaction times did not differ between two-domain and four-domain runs, and so did not depend on the number of rules active in the current run. This is a potentially surprising observation, given classic demonstrations that selection from among a larger set of alternatives is often more difficult (Criss & Shiffrin, 2004; Hick, 1952; Lewis & Anderson, 1976; Schneider & Anderson, 2011). In our data, reaction time is insensitive to the number of alternatives in the active set of tasks, depending only on the relationship between the current and preceding task. In fact, this result was predicted by Oberauer in his theory of procedural working memory (2009), and has now been found several times (Kessler & Meiran, 2010; Souza et al., 2012; van’t Wout et al., 2015). As cognitive load increases, we predict that task information contributing to the overall mental task model is organised into chunks, such that reaction time on a given trial will depend on the complexity of the chunk to which the task belongs, more than the complexity of other chunks.

Insensitivity to the complexity of the current set of tasks, but dependence of the Core DMN response on the overall body of instructions, is reminiscent of the phenomenon of “goal neglect.” Goal neglect describes a situation where someone fails to implement a task rule, despite the rule being understood, remembered, and executable after sufficient prompting (Duncan et al., 1996). Echoing the present results, neglect of a task rule within a complex task structure is insensitive to the number of rules active in any one block, but depends instead on the number of rules that participants are instructed to remember at the beginning of all the blocks (Bhandari & Duncan, 2014; Duncan et al., 2008). The Core DMN’s preferential response to between-domain, relative to within-domain, task switches resembles this “goal-neglect” pattern. While this experiment finds no effect of complexity of the current set of tasks, comparing the results of Zhou et al. (A. X. Zhou et al., 2024; four instructed tasks and no preferential between-domain response) to this and prior experiments (Crittenden et al., 2015; Smith et al., 2018; six or eight instructed tasks and a preferential between-domain response) suggests a dependence on total instructional complexity. Another suggestive link to the phenomenon of goal neglect concerns the observation that rules learned later are more likely to be neglected (Duncan et al., 1996). Here, we find that the Core DMN’s preferential response to between-domain switches emerges for the last-learned group of tasks. These observations suggest that it could be worthwhile for future experiments to examine the relationship between behavioural “goal neglect” and Core DMN involvement more explicitly.

The effect of instruction was also seen in stronger multivariate discrimination of task pairs from different learning groups. The results show that Core DMN is sensitive to abstract cognitive structures established during learning, confirming that teaching the rules in two groups created distinct mental representations. It also implies that Core DMN activation patterns reflect a multi-level task hierarchy (Wen, Duncan, et al., 2020), with increasingly distinct representations for task pairs respectively within a domain, across domains, and across learning groups.

It is, therefore, interesting that this stronger neural discrimination of between-group representations was not accompanied by an increase in Core DMN activity, or by greater behavioural switch costs, at between-group task switches relative to within-group-between-domain switches. Despite multi-level task hierarchies being distinguished in its response pattern, Core DMN activation at task switches appears insensitive to this deeper hierarchy. In turn, this suggests that transition-related DMN activity may only reflect the occurrence of a transition between cognitive chunks, without being scaled by the magnitude of the cognitive or neural pattern difference between these chunks. Of course, it remains possible that different or stronger manipulations of task groupings could reveal task-switch responses with a deeper hierarchy, and it would be useful for future studies to test the generality of the current proposal.

A tentative explanation for the current results is that the later-learned task domains exert strain on the cognitive capacity for representing the overall task model, such that later-learned tasks become chunked more strongly by domain, which then evoke stronger recruitment of the Core DMN at switches between these chunks. Such recruitment of default mode regions could reflect various processes, including detachment from a previously relevant chunk of task rules, or activation of the requirements of the newly appropriate chunk.

This proposal is consistent with the observation that activation of Core DMN is only observed at some task switches, because activation will depend on the chunking structure of the overall set of tasks, which, in turn, depends on the similarity of the tasks, the total complexity of all instructions, and individual capacity limits in maintaining the task model. Integrating results across studies suggests two distinct capacity limits that might be at play. A first capacity limit is the number of tasks or rules that can fit into a “chunk”. The lack of a DMN switch response in traditional two-task designs (Kim et al., 2012), even for tasks belonging to different stimulus domains (Smith et al., 2019), suggests that a single chunk can contain at least two tasks. The emergence of a DMN response when switching among four tasks (A. X. Zhou et al., 2024) suggests that four tasks exceed the capacity of a single large chunk, so they are preferentially represented as four smaller chunks. A second limit is the number of chunks that can be readily maintained. For sets of six or more tasks, the re-emergence of chunking into task pairs (Crittenden et al., 2015; Smith et al., 2018; and the later-learnt domains in the current experiment) suggests that beyond four or five chunks it becomes easier to increase the size of a chunk than to add more chunks. These limits likely depend somewhat on the complexity and confusability of individual rules. This idea relates to previous proposals of multi-level procedural memory (Oberauer, 2009), and would require further exploration, perhaps varying the number of tasks/rules per stimulus domain, as well as the number of domains.

We note here that although Core DMN activity was, on average, significantly greater on switch trials than repeat trials, there was also a substantial degree of variation between people, with a minority of individuals showing the reverse effect. The current study was not powered to investigate the source of this inter-individual variability, but one candidate could be individual differences in working memory capacity (Duncan et al., 2012).

Complexity of the task model is a broad term, which here we have operationalised in two specific ways: the accumulating volume of rules during instruction, and the number of currently relevant tasks among which the participants must switch in a particular run. Previously we have shown that a different form of complexity—the number of stimulus-response mappings within a task—causes activation of multiple demand regions, but deactivation of the DMN (Smith et al., 2019). So far, it appears that the preceding volume of task knowledge as the mental model is assembled may be particularly important in determining the DMN’s response to a task transition. However, other forms of complexity could be considered, such as the depth of the task hierarchy, and the degree of conflict between competing rules. Given the observed effect of learning order, other manipulations of the learning process could be particularly interesting to consider.

Previous work has shown that frontoparietal multiple-demand regions have a phasic response to instruction presentation, which saturates as further instructions are added, and approaches an asymptote after two to three rules (Dumontheil et al., 2011). Although here we did not measure activation during instructions, this gradually saturating frontoparietal response at instruction could plausibly contribute to establishing a task model with well-segregated (less chunked) component rules. This could explain why later-learned tasks collapse into coarser chunks, as the frontoparietal responsiveness asymptotes. Given the role of the hippocampus in learning associations, hippocampal activity at instruction could also influence the effects of learning order. It would be interesting for future studies to examine how responses during instruction relate to the results presented here.

The task-switching effects described here represent an average across all task types. For the univariate fMRI analyses, the between-participant factor coding which particular domains were paired in learning groups showed no significant main effect or interaction, suggesting that results generalised across the particular groupings of tasks. Nonetheless, the present conclusions are limited to the particular set of tasks chosen, and it is possible that Core DMN activity could differ for transitions to or from particular tasks, or when using a different set of tasks. Future research could explore how far the present results generalise to pairings of different task types.

Finally, since the task-switching paradigm involves several cognitive components, such as retrieval of cue-relevant rules, accessing perceptual/semantic content of the stimuli, and inhibition of alternative task rules (Kiesel et al., 2010; Monsell, 2003; Vandierendonck et al., 2010), further research isolating and manipulating cognitive sub-processes may provide further insight into which of them the DMN is sensitive to. While posterior DMN regions have been associated with episodic retrieval (Sestieri et al., 2011; Shapira-Lichter et al., 2013), a recent study suggests that DMN activity is unlikely to reflect retrieval of stimulus-response mappings in tasks of the sort used here (Smith et al., 2019). Nonetheless, the current effect of learning order, and the multivariate discrimination of learning groups, suggests a memory-oriented role of some form. One possibility is that the DMN responses could reflect transitions from a maintenance state to an updating state within working memory (Mayr et al., 2014).

Overall, we document sensitivity of the Core DMN to an intrinsic task hierarchy, suggesting a role in complex cognitive control processes. A better understanding of the role of the DMN during task switching may have relevance to clinical conditions in which DMN activity or connectivity is abnormal, and attention to the appropriate step of a task is impaired. In particular, DMN dysfunction is associated with distractibility in Attention Deficit Hyperactivity Disorder (Fassbender et al., 2009), and with progression of Alzheimer’s disease (Buckner et al., 2008), with poor adherence to complex task plans being a key symptom of both disorders.

The current results also help to reconcile inconsistent observations of DMN activation across task-switch studies, and potentially suggest multiple capacity limits within a mental model of the overall set of tasks. We propose that the Core DMN represents a multi-level task hierarchy, whose chunking structure is determined at instruction, and influences activity at transitions between chunks.

Code for experiment stimuli, data analysis, and figures presented in this manuscript, along with data for generating the figures, can be found here: https://github.com/ashleyxzhou/DMN-TaskSwitch

Unthresholded T-maps of the switch conditions contrasted against task repeat condition can be found here: https://neurovault.org/collections/18470/

A.X.Z., D.J.M., and J.D. designed the research. A.X.Z. collected the data. A.X.Z. and D.J.M. analysed the data. A.X.Z. and D.J.M. wrote the paper. A.X.Z. created the figures. A.X.Z, D.J.M., and J.D. revised the paper. D.J.M. and J.D. supervised the work.

The authors declare no competing interests.

For the purpose of open access, the UKRI-funded authors have applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript arising from this submission. A.X.Z., D.J.M., and J.D. were supported by Medical Research Council Intramural Program MC_UU_00030/7. A.X.Z. was supported by a Gates Cambridge Scholarship.

Supplementary material for this article is available with the online version here: https://doi.org/10.1162/imag_a_00515.