Mind wandering during implicit learning is associated with increased periodic EEG activity and improved extraction of hidden probabilistic patterns

Cognitive control supports goal-directed behavior and performance in attention-demanding, model-based tasks. Conversely, mind wandering (MW)—task-unrelated thoughts that occupy a substantial portion of wakefulness—is often linked to performance costs in sustained attention, executive control, reading comprehension, and explicit sequence learning. Yet MW can sometimes benefit cognition, including creativity and problem-solving, especially when tasks are less attention-demanding or rules/goals are unclear. Probabilistic (statistical) learning is a model-free process enabling incidental extraction of environmental regularities with minimal awareness or focused attention. Prior work suggests MW may not hinder, and might even facilitate, probabilistic learning, but direct evidence with neural correlates is limited. The present study tested whether MW benefits implicit probabilistic learning in the Alternating Serial Reaction Time (ASRT) task, and whether spontaneous versus deliberate MW differentially relates to performance. It also investigated EEG correlates by separating aperiodic and periodic spectral components, and examined whether MW and probabilistic learning share common neural signatures in frequency bands and scalp topography. The overarching hypothesis was that transient, offline-like MW states would be associated with enhanced model-free probabilistic learning and with low-frequency EEG activity indicative of local sleep-like processes, especially during early learning.

Extensive research links MW to poorer performance in attentionally demanding, model-based tasks across laboratory and real-world contexts. However, MW can also aid creative incubation and problem solving. Theoretical perspectives propose that under reduced cognitive control, broadening attentional scope and engaging spontaneous thought may support performance in tasks that are less structured or have unclear rules. Probabilistic learning—automatic acquisition of statistical regularities—often proceeds without explicit awareness or executive control, suggesting it might be resilient to MW’s detrimental effects. Empirical hints support this: lapses of attention can facilitate learning of hidden probabilistic contingencies, and an online ASRT study found improved probabilistic learning but reduced visuomotor accuracy during reported MW. EEG studies associate MW with attenuated sensory processing (reduced evoked potentials) and increased low-frequency activity akin to NREM sleep slow/delta waves; MW’s EEG correlates can be region-specific, resembling local sleep. Increases in theta and alpha power have also been reported during MW. Given these mixed findings and the need to disentangle oscillatory from aperiodic components, parameterizing EEG spectra (e.g., FOOOF) offers a refined approach to relate MW and learning to neural dynamics. The default mode network (DMN) has been implicated in task-unrelated thought and in implicit/statistical learning and automated processing post-learning, suggesting a possible neural bridge between MW and model-free learning.

Participants: Thirty-seven right-handed university students (Mean age 22.1 ± 1.27 years; 30 females) with normal or corrected vision and no psychiatric/neurological/somatic disorders or relevant medications participated for course credit. All provided informed consent; ethics approval from Eötvös Loránd University. One participant had missing data in 10 blocks due to technical issues; mixed models accommodated missingness. Procedure and Task: Participants (12:00–14:00 start) were fitted with a 64-channel EEG cap. They performed a modified ASRT task suited for EEG analyses. Stimuli: central arrow (up/down/left/right) displayed 200 ms, followed by a 500 ms response window (fixation), then a 750 ms fixed ISI (fixation). Errors elicited a 500 ms 'x'; omissions an '!'. The ASRT sequence was an 8-element probabilistic alternating structure (e.g., 2–R–4–R–3–R–1–R), with 24 possible sequences assigned across participants. The task included 2 practice blocks (random) then 30 blocks, each with 80 trials (10 repetitions of the 8-element sequence) preceded by 5 random warm-up trials. High-probability triplets (last element of triplets formed by the alternating structure) occurred 62.5% of trials; low-probability triplets 37.5%. Participants received feedback (mean speed and accuracy) after each block. Post-task awareness checks confirmed no explicit knowledge of the alternating sequence. Thought Probes: After each block, participants rated: Q1) task focus vs MW (1=Not at all focused; 4=Completely), Q2) if not focused, mind wandering vs mind blanking (1=nothing; 4=something in particular), Q3) spontaneity vs deliberateness of attentional state (1=completely spontaneous; 4=completely deliberate). A pre-task quiz ensured understanding. For analyses, Q2 and Q3 were considered only for blocks where Q1 indicated mind wandering (responses 1–2). Behavioral Measures: Trials were categorized as last elements of high- vs low-probability triplets based on preceding two stimuli. Exclusions: trills (e.g., 1-2-1), repetitions (e.g., 2-2-2), first two trials, RT>1000 ms, and incorrect responses (for RT analysis). Two behavioral domains: (1) Probabilistic Learning—difference in accuracy or RT between high- and low-probability triplet last elements per block; (2) Visuomotor Performance—overall accuracy/RT per block regardless of probability. Block-wise and 5-block bin averages were computed. Thought-probe responses were paired to the just-completed block. EEG Recording/Preprocessing: 64-channel BrainAmp system; electrodes per 10% equidistant layout; online reference FCz; re-referenced to mastoids; EOG and chin EMG recorded; impedances <10 kΩ; sampling 1000 Hz; antialiasing and notch filters. Preprocessing in MATLAB/FieldTrip: 0.3–70 Hz band-pass (Butterworth, zero-phase), ICA to remove ocular/muscle/cardiac artifacts (2–4 components), semi-automatic artifact rejection (ft_artifact_zvalue), and visual confirmation. EEG Spectral Decomposition: Artifact-free data segmented into 2 s non-overlapping windows. PSD (1.5–40 Hz; 0.5 Hz resolution) via DPSS multitapers. Spectra parameterized with FOOOF to separate aperiodic (spectral slope and intercept) from periodic peaks. Periodic band power computed on flattened spectra for: Slow (1.5–2 Hz), Delta (2.5–4 Hz), Theta (4.5–8.5 Hz), Alpha (9–13.5 Hz), Beta (14–30 Hz). Bin-wise frequency analyses were also retained. Statistics: Linear Mixed Models (LMMs) tested within-person associations, with random intercepts and, where feasible, random slopes by participant. Predictors were person-mean centered for within-person focus. Models tested: (i) practice effects across blocks on behavior; (ii) MW fluctuations across blocks; (iii) associations between MW and behavior; (iv) associations between EEG measures (aperiodic/periodic, channel-wise) and MW or probabilistic learning, controlling for Block. False discovery rate (FDR) correction applied across electrodes; significance at FDR p<0.05. Given time-dependent effects, EEG-behavior/MW associations were examined separately in the first (Blocks 1–15) and second halves (Blocks 16–30).

Mind wandering trajectory: MW fluctuated within and between individuals (ICC=0.45) and increased over the task (Q1 MW vs Task Focus decreased over blocks; LMM main effect of Block: b = -0.13, 95% CI [-0.19, -0.06], t = -4.19, p<0.001; corroborated by GLMM OR=0.69, 95% CI [0.61, 0.78], p<0.001). Mind blanking vs MW and spontaneous vs deliberate MW did not change significantly over blocks. Learning and performance over practice: Probabilistic Learning (accuracy difference high- minus low-probability) increased with Block (Table 1: b = 0.045, 95% CI [0.016, 0.075], p=0.003); Visuomotor Accuracy (overall accuracy) decreased (b = -0.043, 95% CI [-0.064, -0.021], p<0.001). RT-based measures showed faster responses over practice and faster responses for high vs low probability trials (PL RT: b = 0.09, 95% CI [0.006, 0.17], t=2.11, p=0.0035; Visuomotor RT: b = 0.10, 95% CI [-0.17, -0.02], t=-2.37, p=0.0018). MW-behavior associations (within-person): Greater MW (lower Q1 scores) predicted higher accuracy-based Probabilistic Learning (Table 1: MW vs Task Focus effect b = -1.262, 95% CI [-2.005, -0.519], t = -3.332, p=0.001; lower scores indicate more MW), with a significant interaction indicating stronger benefit earlier in the task (Block × MW: b = 0.071, 95% CI [0.028, 0.111], t=3.390, p=0.001). Splitting the task: MW positively associated with PL in the first half (Blocks 1–15: b = -0.5756, 95% CI [-1.09, -0.06], p=0.029) but not in the second half (b = 0.31, 95% CI [-0.23, 0.84], p=0.26). Additional half-split model showed MW × time interaction (b = 0.88, 95% CI [0.13, 1.63], p<0.02). Visuomotor Accuracy: Focusing on task (higher Q1) predicted better overall accuracy (b = 1.092, 95% CI [0.546, 1.639], t=3.919, p<0.001). MW did not significantly predict RT-based PL or Visuomotor RT. Nature of MW (considering only MW blocks): Spontaneous vs deliberate MW—more spontaneous MW associated with better Probabilistic Learning (estimate = -1.06, 95% CI [-1.68, -0.45], p=0.001) and poorer Visuomotor Accuracy (estimate = 0.81, 95% CI [0.35, 1.26], p<0.001). Mind Blanking vs MW—no significant association with PL; but MW (vs mind blanking) associated with better Visuomotor Accuracy (estimate = 0.44, 95% CI [0.05, 0.83], p=0.028). Interactions with Block were not significant for these subtypes. Interindividual differences: Participants’ mean MW frequency and spontaneity did not predict between-person differences in PL (MW propensity: b = 0.23, p=0.53; spontaneity: b = 0.08, p=0.92). EEG over practice (Block effects): Spectral slope (aperiodic exponent) steepened over time (positive association with Block). Periodic power decreased in Slow and Delta bands and increased in Theta, Alpha, and Beta, with the largest increase in Alpha (8–10 Hz). EEG correlates of MW and PL (controlling Block): First half (Blocks 1–15): Both higher MW and higher PL were associated with increased Slow (1.5–2 Hz) and Delta (2.5–4 Hz) oscillatory activity, peaking over centro-parietal regions, with overlapping topographies. MW also showed widespread increases in Beta power and a slight tendency toward steeper spectral slope at a few frontal/temporal sites; PL showed a slight flattening of slope at one frontal electrode and decreased Beta at right temporo-parietal locations. Second half (Blocks 16–30): MW associated with increased spectral slope (right-lateralized centro-parietal), increased posterior Theta and Alpha, and increased midline Beta; PL showed no significant EEG associations (all FDR-corrected p>0.05).

Findings indicate that MW, despite its documented costs in attention-demanding, model-based tasks, can benefit model-free probabilistic learning. Within individuals, blocks with higher MW showed greater extraction of hidden statistical regularities (especially early in practice), while overall visuomotor accuracy suffered—consistent with a competitive relationship between executive control and model-free learning. The benefit was stronger for spontaneous (vs deliberate) MW, aligning with the idea that spontaneous disengagement reflects reduced executive control that can facilitate incidental learning of environmental regularities. EEG results tie these behavioral patterns to neural dynamics. During early task phases, increased low-frequency periodic activity (Slow and Delta) over centro-parietal areas was associated with both MW and enhanced probabilistic learning, consistent with transient, local sleep-like cortical states that attenuate sensory processing while supporting rapid consolidation/reactivation processes. This converges with accounts of memory consolidation during sleep and waking rest, and suggests that brief offline-like episodes during ongoing task performance may accelerate the formation of predictive representations. Later in practice, MW was linked to a steeper spectral slope and increased posterior Theta/Alpha and midline Beta—patterns resembling the general time-on-task effects—potentially reflecting perceptual decoupling, increased inhibition, and internal attention as visuomotor control becomes more automatized. The lack of PL–EEG associations late in practice suggests distinct phases: early discovery/consolidation of probabilistic structure vs later stabilization not benefited by MW. Overall, the data support a competition framework where reduced cognitive control during MW can shift processing toward model-free mechanisms, aiding probabilistic learning at the cost of immediate visuomotor accuracy. Low-frequency oscillatory activity appears to be a shared neural correlate for both MW and early probabilistic learning, suggesting a mechanistic link via transient offline states and local sleep-like processes.

This study demonstrates that mind wandering can enhance implicit probabilistic learning during task performance, particularly when MW is spontaneous and during early learning phases. Periodic low-frequency (slow and delta) EEG activity over centro-parietal regions was a common neural correlate of both MW and improved probabilistic learning, consistent with transient, local sleep-like states facilitating rapid consolidation and predictive processing. Meanwhile, MW was associated with reduced overall visuomotor accuracy, highlighting domain-specific costs and benefits. Contributions: (1) Replication and extension of MW’s beneficial impact on probabilistic learning in a laboratory EEG setting; (2) dissociation of model-free learning benefits from visuomotor performance costs; (3) identification of shared low-frequency EEG correlates for MW and probabilistic learning; (4) spectral parameterization separating periodic from aperiodic components to clarify neural signatures. Future directions: Use MEG or intracranial recordings to directly test local sleep-like dynamics during MW; experimentally manipulate oscillatory activity or sleep pressure to probe causality; examine DMN involvement in reactivation/predictive coding during MW; increase temporal sampling of MW states with minimal intrusion; characterize MW phenomenology (content, affect, temporal orientation) and its differential impact on learning; diversify samples and balance gender; study phase-specific neural mechanisms of probabilistic learning (discovery vs stabilization).

MW was probed only at block ends, yielding relatively few self-reports and potentially introducing expectancy effects; randomizing probe timing could disrupt probabilistic learning and increase metacognition. Probes were not embedded randomly as in some tasks (e.g., SART), limiting temporal precision. The phenomenology of MW was only coarsely sampled (content presence and spontaneity vs deliberateness); MW is heterogeneous in content, affect, and temporal orientation. The sample was small, young, and female-skewed, limiting generalizability. Although EEG analyses dissociated periodic and aperiodic components, scalp EEG limits inferences about generators; causal roles of oscillations were not tested. One participant had missing blocks, though mixed models accommodate missing data.