Abstract
Mind wandering, occupying 30–50% of our waking time, remains an enigmatic phenomenon in cognitive neuroscience. A large number of studies showed a negative association between mind wandering and attention-demanding (model-based) tasks in both natural settings and laboratory conditions. Mind wandering, however, does not seem to be detrimental for all cognitive domains and was observed to benefit creativity and problem-solving. We examined if mind wandering may facilitate model-free processes, such as probabilistic learning, which relies on the automatic acquisition of statistical regularities with minimal attentional demands. We administered a well-established implicit probabilistic learning task combined with thought probes in healthy adults (N = 37, 30 females). To explore the neural correlates of mind wandering and probabilistic learning, participants were fitted with high-density electroencephalography. Our findings indicate that probabilistic learning was not only immune to periods of mind wandering but was positively associated with it. Spontaneous, as opposed to deliberate mind wandering, was particularly beneficial for extracting the probabilistic patterns hidden in the visual stream. Cortical oscillatory activity in the low-frequency (slow and delta) range, indicative of covert sleep-like states, was associated with both mind wandering and improved probabilistic learning, particularly in the early stages of the task. Given the importance of probabilistic implicit learning in predictive processing, our findings provide novel insights into the potential cognitive benefits of task-unrelated thoughts in addition to shedding light on its neural mechanisms.
Significance Statement
Mind wandering poses an unresolved puzzle for cognitive neuroscience: it is associated with poor performance in various cognitive domains, yet humans spend 30–50% of their waking time mind wandering. We proposed that mind wandering may be beneficial for less attention-demanding cognitive processes requiring automatic, habitual learning. We assessed an implicit probabilistic learning task measuring the ability to extract (without awareness) hidden regularities from the information stream. Participants showed superior performance in probabilistic learning during periods of mind wandering, especially when such task-unrelated thoughts occurred spontaneously without intention. Moreover, mind wandering and probabilistic learning were both associated with slow frequency neural activity, suggesting that mind wandering may reflect a transient, offline state facilitating rapid learning and memory consolidation.
Introduction
Cognitive control is essential for maintaining focus on goal-related information and enabling efficient motor responses, thereby improving performance in a variety of laboratory tasks. Conversely, losing focus and experiencing mind wandering diminishes performance in areas such as sustained attention (Smallwood et al., 2004), executive control (McVay and Kane, 2010, 2012a,b; Andrillon et al., 2021), reading comprehension (Bonifacci et al., 2023), and explicit sequence learning (Brosowsky et al., 2021).
Mind wandering, however, may not be detrimental to behavior in all cases. On the contrary, mind wandering appeared to be beneficial in boosting solutions to a divergent creativity task (Baird et al., 2012) and a convergent problem-solving task (Tan et al., 2015). Broadening the scope of attention and leveraging spontaneous, unconstrained thoughts could be beneficial, especially under cognitively less demanding task conditions or when the goals and rules of the task are not clear (Thompson-Schill et al., 2009; Amer et al., 2016). One example may be statistical (probabilistic) learning, which allows the unintentional extraction of probabilistic regularities of the environment through mere exposure and unsupervised practice (Fiser and Aslin, 2001; Aslin, 2017; Pedraza et al., 2024). Notably, in probabilistic learning tasks, participants automatically acquire stimulus–stimulus dependencies without explicit awareness, cognitive control, or focused attention.
Recent studies suggest that probabilistic learning may not only be unaffected by the disruptive effects of mind wandering but could also derive particular benefits from it. Decker et al. (2023) observed that in a modified flanker task, where distracting items had a hidden probabilistic relationship with the target items, lapses of attention (estimated from response times) facilitated the learning of these apparently goal-irrelevant probabilistic contingencies (Decker et al., 2023). Nonetheless, mind wandering was not directly assessed in this study. The association between mind wandering and probabilistic learning was directly examined by an online study using the Alternating Serial Reaction Time (ASRT) task embedded with thought probes on mind wandering (Vékony et al., 2025). In the ASRT task (Howard and Howard, 1997), implicit statistical learning indicating predictive processes and visuomotor learning are measured simultaneously. Whereas the former reflects the acquisition and anticipation of probabilistic stimulus–stimulus dependencies, the latter captures the efficiency of visuospatial discrimination (Nemeth et al., 2013a). Notably, while participants aim to provide accurate motor responses to a stream of visual inputs, they unintentionally and without awareness extract the probabilistic patterns hidden in the visual stream. This study revealed that participants exhibited improved probabilistic learning but reduced visuomotor accuracy when they reported mind wandering, indicating that mind wandering despite its evident costs may facilitate the extraction of predictable patterns of the environment (Vékony et al., 2025).
The underlying mechanisms of mind wandering, however, are far from being fully understood. Studies on electroencephalographic (EEG) correlates point to attenuated sensory processing during periods of mind wandering, as reflected by changes in evoked potentials and spectral power (Braboszcz and Delorme, 2011; Brandmeyer and Delorme, 2018; Kam et al., 2022). Theoretical speculations suggest that attenuated sensory processing during mind wandering shares some resemblance with states of sleep (Andrillon et al., 2019; Jubera-Garcia et al., 2021; Simor et al., 2023). Moreover, mind wandering may represent a short-lasting, offline state (i.e., transiently decoupled from external inputs) that enhances the processing and consolidation of previously encoded material, resembling the mechanisms of sleep-dependent memory consolidation (Jubera-Garcia et al., 2021; Wamsley, 2022; Vékony et al., 2025). Accordingly, mind wandering has been linked to decreased evoked potentials, indicating diminished sensory processing (Kam et al., 2022), as well as increased, slow-frequency activity reminiscent of the slow and delta waves observed during non-rapid eye movement (NREM) sleep (Braboszcz and Delorme, 2011; Andrillon et al., 2021; Wienke et al., 2021). In addition, sleep-like cortical activity during mind wandering was expressed in a region-specific, topographically localized manner (Andrillon et al., 2021; Wienke et al., 2021), pointing to the links between mind wandering and local sleep (Andrillon et al., 2019). The EEG correlates of mind wandering, however, were not limited to slow frequency activity but also to increases in theta and alpha power (Kam et al., 2022).
Here, we aimed to (1) provide additional support for the benefit of mind wandering in probabilistic learning, (2) explore the neural correlates of mind wandering during task performance focusing on aperiodic and periodic components of the EEG spectra (Donoghue et al., 2020), and (3) examine if mind wandering and probabilistic learning shares neural features in terms of overlapping frequency- and scalp-specific characteristics.
Materials and Methods
Participants
Thirty-seven participants (mean age, 22.1 years ± 1.27, 30 females) were recruited from university students who volunteered for scientific experiments in exchange for partial course credits. Participants had no history of psychiatric, neurological, or chronic somatic disorders; were not currently using medication that could affect alertness, mood, cognition, or sleep; and had normal or corrected-to-normal vision. Only right-handed participants were included in the study as verified by the Edinburgh handedness inventory (Oldfield, 1971). Participants were not informed about the experiment's purpose until after completing the tasks; however, they were debriefed afterward. The study received approval from the Research Ethics Committee of Eötvös Loránd University, and all participants provided informed consent.
Procedure
Participants arrived at the laboratory between 12.00 and 14.00. They were fitted with 64-channel scalp electrodes (see below) and received detailed instructions about the study protocol. They were first introduced to the Alternating Serial Reaction Time (ASRT) task, which involved pressing the key corresponding to the direction of the target stimulus as quickly and accurately as possible, using their left middle and index fingers and their right index and middle fingers. Participants were informed that after each block of the ASRT task, they would be asked three questions to evaluate their thoughts in the previous block and assess their level of mind wandering (MW) which was operationalized as task-unrelated perceptions, thoughts, or memories. A detailed explanation was provided on the different options along with examples of how participants should respond in various scenarios (see Vékony et al., 2025 for more details). Then, participants completed a short quiz to evaluate their understanding of how to answer the questions about their thoughts, with feedback and explanations provided afterward. Participants had the option to retake the quiz or proceed to the task (see Vékony et al., 2025 for more details on the quiz). Following the two initial practice blocks of the ASRT task with random stimuli, participants completed 30 additional blocks of the ASRT task. After responding to the thought probes in each block, participants received performance feedback, which included information on both mean speed and accuracy. To guarantee that the ASRT task functioned in a similar manner to previous studies, we assessed the participants’ conscious knowledge of the hidden sequence following previously validated protocols (Horváth et al., 2022; Vékony et al., 2022). Participants were asked if they noticed anything unusual or any regularities in the task and if so, to elaborate on their response. None of the participants were able to accurately describe the alternating sequence. After completing the ASRT task, participants were asked to respond to demographic questions (age, gender, education, etc.). One participant had missing data due to technical issues in 10 blocks (16–25 blocks), but their remaining data was retained in the analyses.
The ASRT task
The ASRT task embedded with thought probes was displayed on a 24-inch monitor screen in a sound-attenuated room. Responses to the task were recorded using a Cedrus RB-530 response pad (Cedrus Corporation). A modified version of the task suitable for EEG analyses (Kóbor et al., 2018) was used to measure probabilistic learning and visuomotor performance. In this task version, an arrow stimulus appeared at the center of the screen, and participants were instructed to press the corresponding key to the spatial direction (up, down, left, or right) of the arrow as accurately and as fast as they could. The arrow was presented for 200 ms, followed by the presentation of a fixation cross for 500 ms during which the participant could respond. Subsequently, a fixed delay of 750 ms followed, displaying again a fixation cross, and then, the next trial appeared (applying fixed inter stimulus intervals). In case of a wrong response, a “×” appeared in the middle of the screen for 500 ms, and in case of a missing response, an “!” appeared for the same duration. Following this, a fixation cross appeared for 250 ms before the next trial began. The stimuli followed a probabilistic eight-element sequence, alternating between patterned and random elements (e.g., 2 – R – 4 – R – 3 – R – 1 – R, where “R” indicates a random direction and the numbers represent predetermined directions from up, down, left, and right). Each participant was assigned one of 24 possible sequences, which they encountered throughout the task. The ASRT task comprised 30 blocks, each consisting of 10 repetitions of the eight-element sequence (80 trials) preceded by five random trials at the beginning of each block for warm-up purposes. Following each block, participants took a brief break and were instructed to respond to the mind wandering thought probes before resuming the task (Fig. 1B). The ASRT task incorporated a probabilistic sequence structure where certain runs of three consecutive stimuli (triplets) appeared with a higher probability (high-probability triplets) than others (low-probability triplets). A trial refers to a single element in the sequence, which could be either a pattern or random element and, importantly, also the last element in a high- or low-probability triplet. It is crucial to note that the analysis focuses on whether the provided trial constitutes the final element of a high- or low-probability triplet, rather than its classification as a pattern or random element within the alternating sequence. For instance, in a sequence like 2 – R – 4 – R – 3 – R – 1 – R, triplets such as 2-X-4, 4-X-3, 3-X-1, and 1-X-2 (where X represents the middle element of a triplet) occurred more frequently than triplets like 2-X-1 or 2-X-3 (Fig. 1C). It is important to highlight that when referring to triplet type later on, the focus is on trials serving as the final element of a high- or low-probability triplet. Throughout the task, a total of 64 distinct triplets could potentially occur (16 with high probability and 48 with low probability). High-probability triplets could be formed either by having two pattern trials and one random trial in the center (occurring in 50% of trials) or by having two random trials and one pattern trial in the center (occurring in 12.5% of trials). Among all trials, 62.5% represented the last element of a high-probability triplet, while 37.5% were assigned to the last element of a low-probability triplet (Fig. 1D).
Experimental design and task structure of the ASRT task. A, In the ASRT task, participants had to press keys corresponding to the direction of an arrow that appeared in the center of the screen. Every second trial was part of an 8-element probabilistic sequence. Random elements were inserted among pattern elements to form the sequence (e.g., 2-R-4-R-3-1-R, where numbers indicate the direction of pattern trials (up, down, left, or right), and r represents random directions of the four possible directions. B, The experiment consisted of 30 blocks with mind wandering thought probes administered after each block of 80 trials. Participants were asked to reflect on their thoughts during the block and decide whether the last block was dominated by (1) mind wandering versus task focus, (2) mind wandering versus mind blanking, and (3) spontaneous versus deliberate mind wandering. C, Formation of triplets in the task. Pattern elements are represented by red backgrounds (they are constantly pointing at that direction throughout the task), and random elements are represented by blue backgrounds (they are always chosen from the four possible directions randomly). Every trial was categorized as the third element of three consecutive trials (a triplet). It is worth highlighting that the analysis depends on whether the given trial is the last element of a high- or low-probability triplet, rather than its classification as a pattern or random element within the alternating sequence. The probabilistic sequence structure resulted in a higher occurrence of some triplets (high-probability triplets) than others (low-probability triplets). Please note that only three triplets are highlighted on this subfigure for visualization purposes [2(P)-1(R)-4(P) as a pattern-ending high-probability triplet, 2(R)-3(P)-4(R) as a random-ending high-probability triplet, and 4(R)-1(P)-2(R) as a random-ending low-probability triplet]. However, every three consecutive elements form either a high- or low-probability triplet. Therefore, in the above example, from these eight consecutive elements—2(P)-1(R)-4(P)-2(R)-3(P)-4(R)-1(P)-2(R)—six triplets can be formed. If we consider the pattern element 2 as a starting point, then the triplets are in the following order: 2(P)-1(R)-4(P), 1(R)-4(P)-2(R), 4(P)-2(R)-3(P), 2(R)-3(P)-4(R), 3(P)-4(R)-1(P) (these are all high-probability triplets), and 4(R)-1(P)-2(R) (low-probability triplet). D, The formation of high-probability triplets could have involved the occurrence of either two pattern trials and one random trial at the center, which transpired in 50% of trials, or two random trials and one pattern trial at the center, which occurred in 12.5% of trials. In total, 62.5% of all trials constituted the final element of a high-probability triplet, while the remaining 37.5% were the final elements of a low-probability triplet.
Mind wandering thought probes
Following each block of the ASRT task, participants were prompted to reflect on their thoughts during the last (just completed) block and respond to questions regarding their attentional states (Fig. 1B). More specifically, they were asked to rate if in the last block (1) they engaged in mind wandering or maintained task focus; (2) in case they experienced mind wandering, whether they were thinking of something particular, or did not think about anything (“mind blanking”); and (3) whether their attentional state was controlled deliberately or was rather spontaneous. This way, after each block, participants were presented with three items to be rated on a four-point scale: (Q1) To what degree were you focusing on the task just before this question? (1, Not at all; 4, Completely), (Q2) To the degree to which you were not focusing on the task, were you thinking of something in particular or just thinking about nothing? (1, I was thinking about nothing; 4, I was thinking about something in particular), and (Q3) Were you deliberate about where you focused your attention (either on-task or elsewhere) or did it happen spontaneously? (1, I was completely spontaneous; 4, I was completely deliberate). While the first question has been used in prior studies (Alexandersen et al., 2022; Groot et al., 2022), the latter two were tailored to differentiate between MW with reportable content versus MB (Q2) and between spontaneous and deliberate MW (Q3). Although it is common to directly distinguish between on-task periods and either MW versus MB, or unintentional versus intentional MW with specific questionnaires (Seli et al., 2018), we chose to explore these aspects of MW in two follow-up questions to avoid overwhelming participants with numerous response options and to gain a more nuanced understanding of their mental states during the ASRT task (Aasen et al., 2024; Drevland et al., 2024).
EEG recording and preprocessing
EEG activity was measured by a 64-channel recording system (BrainAmp amplifier and BrainVision Recorder software, Brain Products). The Ag/AgCl sintered ring electrodes were mounted in an electrode cap (EasyCap) on the scalp according to the 10% equidistant system. During acquisition, electrodes were referenced to the FCz electrode but were rereferenced to the average of the mastoids before further data processing. Horizontal and vertical eye movements were monitored by EOG channels. EMG electrodes to record muscle activity were placed on the chin. All electrode contact impedances were kept below 10 kΩ. EEG data was recorded with a sampling rate of 1,000 Hz. Antialiasing digital filters and notch filters to remove power line noise were applied. Preprocessing of EEG data was performed by using custom-made scripts in MATLAB (version 9.14.0.2206163, R2023a, The MathWorks) as well as functions of the Fieldtrip toolbox (Oostenveld et al., 2011). EEG recordings were bandpass filtered between 0.3 and 70 Hz (with Butterworth, zero phase forward and reverse digital filters). We performed independent component analysis (ICA) to identify cardiac, eye movement, and other muscular artifacts using FieldTrip routines (Oostenveld et al., 2011). Independent components, primarily two to three (up to four maximum), representing eye movements and muscular artifacts, were semiautomatically detected and identified by inspecting waveforms and their topographical distribution (Viola et al., 2009). Additionally, a semiautomatic artifact rejection tool was applied using the ft_artifact_zvalue function of the FieldTrip routine. This procedure involves filtering the data and averaging it across channels after z-transformation. Accumulated z-scores are then visualized, and thresholds are flexibly set and adapted to each recording. Subsequently, trials identified as artifacts were visually inspected to confirm and discard those with technical or movement-related artifacts.
Data analyses I: behavioral and thought probes
We used MATLAB to process behavioral data assessed in the ASRT task. Each trial (representing an arrow pointing to one of the four directions) was categorized based on the two preceding trials as the last element of a high- or low-probability triplet. Trills (repetitions where the first and the third element of the triplet is the same with a different trial in between, such as 1-2-1, 2-1-2, 3-1-3, 1-4-1, etc.), repetitions (e.g., 2-2-2), the first two trials, and trials with a reaction time above 1,000 ms were removed from the analysis. Inaccurate responses were also removed from the analyses of RTs. We defined two types of behavioral measures: Probabilistic Learning and Visuomotor Performance. Probabilistic Learning was operationalized as the difference in RT/accuracy between high-probability and low-probability trials (i.e., between the third element of a high-probability triplet and the third element of a low-probability triplet). Visuomotor Performance, on the other hand, was operationalized as the overall RT/accuracy on the task and its changes over time (i.e., lower accuracy in later blocks) regardless of the trial probability. Probabilistic Learning and Visuomotor Performance scores were extracted considering the mean accuracy scores (percentage of correct responses/all trials) and the median of RTs of accurately responded trials and for each block. Similarly, the same measures were extracted for each time bin comprising the averages for chunks of five consecutive blocks (Block 1–5, 6–11…25–30). Since mind wandering was previously linked to accuracy-based measures of probabilistic learning (Vékony et al., 2025), we focused on this metric, but also examined RT-based measures. Responses to mind wandering thought probes were similarly extracted for each block and time bin. For time bins, responses to thought probes were averaged for each chunk of five consecutive blocks. Scores reflecting Mind Wandering versus Task Focus were present for each block. However, scores representing Mind Wandering versus Mind Blanking, and Spontaneous versus Deliberate mind wandering were only considered when participants reported a tendency to mind wander (responses 1 and 2) instead of focusing on the task (responses 3 and 4) during the block (i.e., to Q1). Responses to thought probes were paired with behavioral performance in the same blocks, that is, mind wandering reports were linked always with behavioral performance in the blocks that the participants just finished.
Data analyses II: EEG analyses
EEG analyses were performed by using custom-made scripts and Fieldtrip routines in MATLAB. We extracted the aperiodic and periodic spectral components of EEG activity measured during the 30 blocks of the ASRT task. To quantify the spectral slope as well as the oscillatory (periodic) activity of the EEG signal, we applied the FOOOF (fitting oscillations and one over f) method (Donoghue et al., 2020). Separating the EEG spectra into aperiodic and periodic components is an efficient technique to delineate if changes in power are caused by a shift in aperiodic background activity following a power law or increases/decreases in oscillatory activity beyond the aperiodic background. Moreover, the parametrization of aperiodic and periodic activity reduces the redundancy of the power spectrum featuring largely correlated power values across different frequencies (Donoghue et al., 2020; Bódizs et al., 2021; Gerster et al., 2022; Schneider et al., 2022). Parametrization of the EEG spectra proved to be a sensitive approach to track different states of consciousness (Colombo et al., 2019; Bódizs et al., 2021), cognitive performance (Ouyang et al., 2020), and shifts in modality-specific attention (Waschke et al., 2021).
Artifact-free EEG data segmented into 2-s-long nonoverlapping time windows was subjected to fast Fourier transformation (FFT) using discrete prolate spheroidal sequences (DPSS) multitapers to attenuate spectral leakage. Power spectral densities (PSDs) between 1.5 and 40 Hz with 0.5 Hz resolution were obtained and were averaged over the segments within each block and each participant. Hence, the power spectra were used as the input to extract the aperiodic and periodic components. In brief, the FOOOF method computes a linear fit to the spectrum represented in a log-log scale. This linear fit is then subtracted from the spectrum yielding a flattened spectrum. Using the flattened spectrum a relative and an absolute threshold is set to two times the standard deviation of the flattened spectrum. Next, the method fits Gaussian functions to the largest peak exceeding the threshold, subtracts it from the spectrum, and then iteratively proceeds to the next largest peak until no more peaks surpassing the threshold are detected. The oscillatory components are then derived by fitting a multivariate Gaussian to all extracted peaks simultaneously. Following these iterations, the initial fit is reintegrated into the peak-free flattened PSD, resulting in the aperiodic component of the PSD. Subsequently, this aperiodic component undergoes another fitting process, producing the final fit with parameters for y-intercept and slope. We used the spectral slope in our analyses, which is assumed to reflect neural inhibitory and excitatory balance (Gao et al., 2017; Donoghue et al., 2020), with larger values (steeper slopes) reflecting stronger inhibition. Periodic activity was inferred from the flattened PSD. Band-wise oscillatory activity was obtained by averaging power values between 1.5–2 Hz (slow), 2.5–4 Hz (delta), 4.5–8.5 Hz (theta), 9–13.5 Hz (alpha), and 14–30 Hz (beta), but bin-wise values were also retained for further analyses.
Statistical analyses
Statistical analyses and data visualizations were performed in Rstudio (version 1.4 1717; Allaire, 2012) and MATLAB. Since multiple block-wise values for behavioral, self-report, and EEG variables were nested within each individual, linear mixed models (LMMs) were employed to examine within-person associations between our variables of interest allowing for random effects (random intercept and slope) by participants. For each LMM, we fitted random intercepts and random slopes for within-participant predictors. In case of convergence issues (e.g., singular fit), a simpler model was examined, that is, random slopes were removed, and random intercepts were fitted only. Model fits were compared by the log-likelihood (LL) and the Akaike information criterion (AIC) and the best-fitting models were considered. To focus on within-person associations, the numerical predictors were mean-centered (to the individual's mean values; Wang and Maxwell, 2015). The relevant subsections of the Results section provide detailed descriptions of the specific outcome variables, as well as the fixed and random effects to ease the understanding of the specific models. In brief, separate LMMs were used to (1) examine the rate of Probabilistic Learning and Visuomotor Performance as a function of practice throughout the successive blocks; (2) examine within-person fluctuations in self-reported mind wandering throughout the successive blocks; (3) examine within-person associations between self-reports of mind wandering and behavioral performance (Probabilistic Learning and Visuomotor Performance) in the ASRT task; and (4) examine the association of block-wise within-person fluctuations in aperiodic and periodic EEG activity with mind wandering and probabilistic learning. Separate LMMs were run to test the associations of EEG measures (outcome variables) with mind wandering and probabilistic learning (predictors in separate models). In order to control the confounding effect of time on EEG activity, Block was entered as a continuous, mean-centered predictor in addition to probe responses about attentional focus (mind wandering vs task focus), as well as in addition to probabilistic learning in the LMMs with EEG measures as outcome variables. To explore the region-specific, topographical aspects of these associations, LMMs were performed relating mind wandering and probabilistic learning with EEG measures at each electrode location. To address the issue of multiple comparisons, p values were adjusted with FDR (false discovery rate) correction (Benjamini and Hochberg, 1995). FDR-corrected p values below 0.05 were considered significant. The data of all participants were included in the study. One participant had missing data in some of the blocks due to technical issues, but their remaining data was included in the analyses since mixed models can properly handle missing data.
Results
The tendency to mind wander fluctuates and gradually increases throughout the task
First, we examined whether the tendency of mind wandering changed throughout the task. Inter- and intraindividual variability and the relative frequencies of Mind Wandering versus Task Focus self-reports are depicted in Figure 2. Mind wandering was highly variable across the successive blocks showing considerable inter- and intraindividual variability (ICC = 0.45; Fig. 2A). Participants tended to mind wander more at later stages of the task (Fig. 2B,C). This was revealed by a LMM (with random intercepts and slopes by participant) regressing self-reports of Mind Wandering versus Task Focus (outcome variable on a 4-point scale for thought probe Q1) on Block number (predictor), as indicated by the main effect of Block (b = −0.13, 95% CI = [−0.19, −0.06], t = −4.19, p < 0.001). The model showed better fit than the null model which included the random intercept of participant only but excluded the predictor Block: differences in LL and AIC compared with the null model were 71.5 and −137, respectively. Likewise, a GLMM regressing dichotomized Mind Wandering versus Task Focus scores (binary outcome variable) showed a similar main effect of Block (odds ratio = 0.69, 95% CI = [0.61, 0.78], t = −6.3, p < 0.001). The model showed better fit than the null model which included the random intercept of participant only: differences in LL and AIC compared with the null model were 20.1 and −38.2, respectively. To study if the phenomenological aspects of MW experiences varied throughout the task, we also examined if Mind Blanking versus Mind Wandering and Spontaneous versus Deliberate MW were significantly predicted by Block. Block number as a predictor in LMM (with random intercepts and slopes by participant) did not significantly predict the tendency of Mind Blanking versus Mind Wandering (b = 0.04, 95% CI = [−0.13, 0.22], t = 0.47, p = 0.64; model fit showed negligible differences compared with the null model: differences in LL and AIC compared with the null model were −0.1 and 1.9, respectively) or Spontaneous versus Deliberate MW (b = −0.03, 95% CI = [−0.17, 0.11], t = −0.44, p = 0.66; model fit showed negligible differences compared with the null model: differences in LL and AIC compared with the null model were 1.65 and −1.3, respectively).
Variability and temporal trajectory of mind wandering during the ASRT task. A, Mind wandering exhibited considerable intraindividual variability throughout the successive blocks. Each line (distinguished by different colors) represents different participant's responses (MW vs Task Focus) over the blocks. Values on the y-axis are centered for visualization only. The plot shows that mental experience, as assessed by thought probes, fluctuates between mind wandering and task focus throughout the blocks. B, Mind Wandering versus Task Focus scores over the successive blocks. The plot indicates that self-reports of mind wandering are gradually more prevalent as the task progresses. Values were centered block-wise to the overall individual means for visualization purposes. Lower scores on the y-axis indicate increased mind wandering. The red points represent the participant's (centered) responses in each block; the gray boxplots represent the interquartile range, the medians (black lines), and the Tukey's fences (with whiskers showing the 1.5 interquartile ranges). The red line represents the linear regression model fitted to the data points. C, The amount of Mind Wandering versus Task Focus states by dichotomized values. Besides the final block, mind wandering becomes more prevalent over the course of the task.
Mind wandering is linked to better probabilistic learning but poorer visuomotor accuracy
To investigate if mind wandering was associated with acquiring hidden probabilistic regularities and with overall visuomotor performance, we entered the predictors Block, Mind Wandering versus Task Focus scores, as well as their interaction, on the outcome measures of (accuracy and RT based) Probabilistic Learning and Visuomotor Performance (termed as Visuomotor Accuracy and Visuomotor RT), in separate LMMs. (Notably, reports of mind wandering always referred to the block that the participants just completed.) With regard to accuracy, Probabilistic Learning was defined as the difference in block-wise responses to high- versus low-probability trials, whereas Visuomotor Accuracy represented overall, block-wise accuracy regardless of trial types. First, we evaluated if learning occurred over the course of the task. As shown in Table 1, a significant main effect of Block emerged indicating increased Probabilistic Learning over the course of the task, that is, participants appeared to be gradually better in extracting the statistical pattern hidden in the task, becoming more accurate (Fig. 3A,C) in responding to high-probability versus low-probability trials. Block was also significantly and negatively associated with Visuomotor Accuracy, pointing to overall less accurate (Fig. 3B,D) responses as the task progressed, regardless of the probabilistic nature of trials. Reaction time-based measures exhibited a similar learning pattern. LMM random intercept models with RT-based indices of Probabilistic Learning and Visuomotor RT as outcome variables showed a significant main effect of Block, pointing to gradually better probabilistic learning performance (b = 0.09, 95% CI = [0.006, 0.17], t = 2.11, p = 0.0035) and improved visuomotor performance (b = −0.10, 95% CI = [−0.17, −0.02], t = −2.37, p = 0.0018). That is, participants were faster in responding to high- versus low-probability trials and showed generally faster reaction times (regardless of trial types) as the task progressed.
Probabilistic Learning and Visuomotor Accuracy over successive blocks and bins. Probabilistic Learning is quantified by the difference in accuracy (A, C; percentage of correct responses) between high-probability and low-probability trials averaged within each block. Probabilistic Learning gradually increases throughout the task due to relatively more accurate responses to high- versus low-probability trials. On the other hand, Visuomotor Accuracy (overall accuracy regardless of trial types) (B, D) is gradually reduced as the task progresses, reflecting less accurate responses regardless of trial types. The upper graphs depict performance across each block. In the lower graphs (to facilitate visualization), the same performance is averaged over successive steps of five consecutive blocks (Blocks: 1–5, 6–10…26–30) marked as bins.
Mind wandering and ASRT task performance
As shown in Table 1, accuracy-based indices of Probabilistic Learning were significantly higher during periods when participants reported a tendency to mind wander versus focusing on the task (b = −1.26, 95% CI = [−2.00, −0.52], t = −3.33, p = 0.001; Fig. 4A). Moreover, the interaction between Mind Wandering versus Task Focus with Block (b = 0.07, 95% CI = [0.03 0.11], t = 3.39, p = 0.001) suggested that mind wandering was associated with better probabilistic learning at earlier phases of the task. Model selection revealed that the most complex model including the interaction term had the best fit compared with the simpler main effects model and compared with the model which included the main effect of the block only. In a following step, we aimed to parse out the interaction between mind wandering and the time spent with the task. Since in the ASRT task the performance fluctuates considerably from block to block, we divided the responses into two halves (15 blocks each). This division provided a reasonable compromise, balancing an adequate number of sampling points (for mind wandering) with the opportunity to examine how time influences the association between MW and probabilistic learning. Therefore, to explore this interaction in more detail, and to examine if mind wandering exhibits a distinct association with probabilistic learning in the first (blocks 1–15) versus the second (blocks: 16–30) half of the ASRT task, we employed a further LMM with Mind Wandering versus Task Focus and ASRT time [first half (blocks 1–15) vs second half (blocks 16–30)] as predictors of the accuracy-based measure of Probabilistic Learning. We observed a significant interaction between Mind Wandering and ASRT time (first vs second part of the ASRT task; b = 0.88, 95% CI = [0.13 1.63], t = 2.31, p < 0.02), in addition to the significant main effects of ASRT time (b = 0.77, 95% CI = [0.25 1.28], t = 2.94, p < 0.003) and Mind Wandering versus Task Focus (b = −0.62, 95% CI = [−1.16 −0.09], t = −2.3, p < 0.02) pointing to increased Probabilistic Learning when participants experienced more mind wandering, specifically in the first part of the ASRT task. The best-fitting model proved to be the more complex model including the interaction term which outperformed the simpler model including only the two main effects (LL and AIC differences were 2.7 and −4, respectively). Post hoc contrasts showed that the associations between mind wandering and probabilistic learning between the first (EMM: 0.71 95% CI [0.21 1.20]) and the second part of the ASRT task (EMM: 1.42 95% CI [0.97 1.97]) were significantly different (estimate: −0.8, t-ratio = −2.9 p < 0.003), and a further analyses splitting the data into two halves indicated that the significant positive association between mind wandering and probabilistic learning was apparent in the first half (b = −0.5756, 95% CI = [−1.09 −0.06], t = −2.19, p = 0.029) but was not significant in the second half of the ASRT task (b = 0.31, 95% CI = [−0.23 0.84], t = 1.13, p = 0.26).
Mind wandering is associated with enhanced Probabilistic Learning and attenuated Visuomotor Accuracy. The results of random intercept and fixed slope LMMs are visualized regressing block-wise centered Mind Wandering versus Task Focus scores on (accuracy-based) Probabilistic Learning (A) and Visuomotor Accuracy (B). Lower values on the x-axes (A, B) indicate increased mind wandering. Similarly, the results of random intercept and fixed slope LMMs are visualized regressing block-wise centered Spontaneous versus Deliberate Mind Wandering on (accuracy-based) Probabilistic Learning (C) and Visuomotor Accuracy (D). Lower values on the x-axes (C, D) indicate increased spontaneous mind wandering. Colored lines represent random intercepts and fixed regression slopes for each participant. Points depict values for the variables of interest of all participants at each block.
With regard to Visuomotor Accuracy, Mind Wandering versus Task Focus was positively associated with overall accuracy (b = 1.09, 95% CI = [0.54 1.63], t = 3.91, p < 0.001), indicating that participants committed generally less errors when they tended to focus more on the task (Fig. 4B). The interaction between Block and Mind Wandering versus Task Focus was not significant (b = −0.03, 95% CI = [−0.05 0.00], t = −1.7, p = 0.08). Model selection revealed that the most complex model including the interaction term had the best fit compared with the simpler main effects model and compared with the model which included the main effect of the block only. Reaction time-based outcome measures of Probabilistic Learning and of Visuomotor Accuracy were not significantly predicted by self-reports of Mind Wandering (for Probabilistic Learning: b = 0.69, 95% CI = [−1.48 2.88], t = 0.62, p = 0.53; for Visuomotor Accuracy: b = 0.02, 95% CI = [−1.99 2.03], t = 0.02, p = 0.98) or the interaction between Mind Wandering and Block (for Probabilistic Learning: b = 0.03, 95% CI = [−0.09 0.14], t = 0.47, p = 0.64; for Visuomotor Accuracy: b = 0.01, 95% CI = [−0.10 0.11], t = 0.12, p = 0.91).
The nature of mind wandering relates to probabilistic learning and visuomotor accuracy
We examined if the type of mind wandering was associated with behavioral performance. Therefore, in our next analyses, we considered only those blocks when participants reported to experience mind wandering during the task. Mind Blanking versus Mind Wandering as well as self-reports of Spontaneous versus Deliberate mind wandering (MW) were used as mean-centered predictors of our behavioral measures. Since mind wandering was associated with accuracy-based measures of learning, we only considered accuracy-based Probabilistic Learning and Visuomotor Accuracy as our outcome variables. The predictors—Block, the nature of mind wandering, and their interaction—were evaluated in separate LMMs based on model fit indices. We started with the most complex interaction models (Block × Mind Blanking vs Mind Wandering and Block × Spontaneous vs Deliberate MW) and compared them with their simpler counterparts (e.g., Block + Mind Blanking vs Mind Wandering), progressively simplifying until identifying the best-fitting model. Neither the experience of Mind Blanking versus Mind Wandering nor its interaction with Block was significantly associated with Probabilistic Learning. Model comparisons revealed that the simplest null model had the best fit as the predictor Block was not significant in this set of analyses either. On the other hand, mind wandering reports in contrast to mind blanking were significantly associated with better Visuomotor Accuracy (see Table 2 for more details). The best-fitting model included the main effects of Block and Mind Blanking versus Mind Wandering but excluded the interaction term (Table 2). Notably, more spontaneous (vs deliberate) mind wandering was significantly associated with better Probabilistic Learning (Fig. 4C) and poorer Visuomotor Accuracy (Fig. 4D; see Table 2 for more details). The interactions between block number and the nature of mind wandering were not significant on either of the outcome behavioral measures, and the best-fitting models included the main effects of Block and Spontaneous versus Deliberate MW but excluded the interaction term. Table 2 summarizes the statistical parameters of the best-fitting models of these analyses.
Probabilistic learning and visuomotor accuracy predicted by the subtypes of mind wandering
Individual differences in mind wandering are not significantly associated with probabilistic learning
To examine whether participants who mind wandered more frequently during the task demonstrated better probabilistic learning, we included individual averages of mind wandering scores (the individual averages of the 30 blocks) in models predicting probabilistic learning. This analysis aimed to explore whether interindividual differences in mind wandering contribute to variability in probabilistic learning, beyond the associations observed at the intraindividual level. Interindividual differences in mind wandering propensity were not significantly associated with Probabilistic Learning (b = 0.23, 95% CI = [−0.48 0.92], t = 0.63, p = 0.53). Likewise, interindividual differences in spontaneous versus deliberate mind wandering were not significantly associated with Probabilistic Learning (b = 0.08, 95% CI = [−1.48 1.62], t = 0.10, p = 0.92).
Slow periodic EEG activity is significantly associated with both mind wandering and probabilistic learning
To examine the neural correlates of mind wandering and behavioral performance, we extracted the aperiodic component (the spectral exponent) and the periodic EEG components as measured during the blocks when participants actively engaged with the task. The spectral exponent and the periodic band-wise power at each channel served as outcome variables in separate LMMs. First, we explored whether time as reflected by Block number was associated with changes in EEG activity; therefore, aperiodic and periodic EEG measures at each location were predicted by Block in separate LMMs. As indicated in Figure 5, Block showed a considerable association with both aperiodic and periodic EEG activity. The slope of the EEG spectra showed a positive association with Block indicating a steepening of the EEG spectral slope as the task progressed. With regard to the periodic components, oscillatory activity in the slow and delta frequency range decreased, whereas activity in the theta, alpha, and beta ranges increased as the task progressed. The most robust increase was evidenced in the alpha range.
Aperiodic and periodic EEG activity predicted by Block number. LMMs predicting channel-wise EEG features by Block number indicate a robust change in EEG measures as the task progresses. The top graph highlights the t values of individual LMMs predicting oscillatory (periodic) EEG activity in each channel and frequency bin by Block. The bottom headplots indicate the t statistics of LMMs with aperiodic (Slope) and periodic activity (the later averaged over frequency bands) as outcome variables and Block as the predictor. Asterisks on the headplots indicate statistically significant models after FDR correction. The findings indicate steepening spectral slope, decreasing slow and delta and increasing theta, alpha, and beta oscillatory activity as the task progresses. The most pronounced increase occurred in the alpha (8–10 Hz) frequency range as indicated by the bin-wise analyses (top graph).
Next, we examined the associations of Mind Wandering and Probabilistic Learning with EEG activity. Due to the robust effect of block number on EEG measures, we controlled for this confounding factor and hence, performed LMMs predicting the spectral slope as well as the band-wise periodic EEG components by the fixed effects factors Block and Mind Wandering and by Block and Probabilistic Learning in separate models at each electrode location (including by-participant random effects). Notably, our behavioral analyses reported above indicated a distinct association between mind wandering and probabilistic learning as a function of time. Given that the positive association between mind wandering and probabilistic learning was evidenced in the first part of the task but was not apparent in the second part, we analyzed the first (1–15 blocks) and second half (16–30 blocks) of the task periods in separate models. Channel-wise analyses showed that in the first half of the task (1–15 blocks), periodic activity in the slow and delta frequency range was positively associated with both increased Mind Wandering and Probabilistic Learning peaking at centroparietal sites and to some extent at frontal and frontolateral locations. The association of slow and delta oscillatory activity with Mind Wandering and Probabilistic Learning showed considerable topographical overlaps (Fig. 6). On the other hand, higher Mind Wandering was associated with a slight tendency (affecting only four electrode sites at frontopolar and right temporal locations) of a steepening spectral slope and with a global increase in periodic activity in the beta range (comprising frontal, central, and parietal sites), while increased Probabilistic Learning was associated with a tendency of a flattening slope (present in only one frontal electrode) and by decreased beta activity at right temporoparietal locations. Significant EEG correlates of mind wandering in the second half (16–30 blocks) of the task consisted of a right lateralized centroparietal increase in the spectral slope, and widespread increases in theta and alpha oscillatory activity over posterior regions, as well as significant increase in beta oscillatory activity peaking over midline locations along the anterioposterior axis. Probabilistic learning showed no significant associations in any of the electrode sites (all FDR-corrected p values >0.05) with aperiodic or periodic EEG activity in the second half of the task (Fig. 7).
Topographical aspects of the associations of Mind Wandering and Probabilistic Learning with EEG activity in the first part (1–15 Block) of the ASRT task. LMMs predicting channel-wise EEG features by the independent variables Block and Mind Wandering as well as by Block and Probabilistic Learning were performed including the data of the first half of the task. Mind Wandering was associated with widespread oscillatory activity in the slow, delta, and beta frequency ranges, whereas Probabilistic Learning was linked to increased oscillatory activity in the slow and delta range. The t statistics (t values) of the predictors Mind Wandering (headplots on the left side, and magenta-colored lineplots in the middle) and Probabilistic Learning (headplots on the right side and green-colored lineplots in the middle) are shown for each EEG feature and channel. The y-axes of the magenta-and green-colored lineplots correspond to the color-coded headplots (both showing the t values), on the left (MW), and the right (Prob.Learning), respectively. Significant channels after FDR correction are highlighted by black points in the headplots and by black (MW) and blue point (Prob.Learning) markers in the lineplots in the middle. Mind Wandering scores were inverted to facilitate understanding, that is, higher scores reflected higher mind wandering. Slow and delta oscillatory correlates of Mind Wandering and Probabilistic Learning showed considerable overlaps, especially in centroparietal sites.
Topographical aspects of the associations of Mind Wandering and Probabilistic Learning with EEG activity in the second part (16–30 Block) of the ASRT task. LMMs predicting channel-wise EEG features predicted by Block and Mind Wandering as well as by Block and Probabilistic Learning were performed including the data of the second half of the task. Mind Wandering was associated with a steeper spectral slope over right lateralized centroparietal sites and widespread increases in posterior theta and alpha and midline beta oscillatory activity. Probabilistic Learning was not associated with changes in EEG features in the second half of the task. The t statistics (t values) of the predictors Mind Wandering (headplots on the left side and magenta-colored lineplots in the middle) and Probabilistic Learning (headplots on the right side and green-colored lineplots in the middle) are shown for each EEG feature and channel. The y-axes of the magenta- and green-colored lineplots correspond to the color-coded headplots (both showing the t-values), on the left (MW), and the right (Prob.Learning), respectively. Significant channels after FDR correction are highlighted by black points in the headplots and by black (MW) markers in the lineplots in the middle. Mind wandering scores were inverted to facilitate understanding, that is, higher scores reflected higher mind wandering.
Discussion
A wide range of studies point to the harmful impact of mind wandering on everyday activities in natural settings (McVay et al., 2009; Unsworth et al., 2012; Szpunar et al., 2013; Yanko and Spalek, 2014) and on cognitive processes measured in laboratory conditions (Mooneyham and Schooler, 2013). Such instances of task-unrelated thoughts often lead to erroneous responses, failure to detect targets and discard distracting items, or a reduced ability to accurately encode, store, or recall information (Miller, 2000; Braver, 2012; McVay and Kane, 2012b; Amer et al., 2016; Gratton et al., 2018; Andrillon et al., 2021; Blondé et al., 2022). Notably, the high prevalence of reports on the negative effects of mind wandering is due to studies focusing on attention-demanding cognitive operations, often defined in computational terms as model-based processes (Shohamy and Daw, 2014). Whereas model-based computations are cognitively more demanding, require executive control, and effortful, focused attention, model-free learning relies on acquiring less attention-demanding dependencies between stimuli (Daw et al., 2005; Otto et al., 2013; Pedraza et al., 2024). Here we showed that engaging in mind wandering might facilitate model-free processes, as exemplified by probabilistic learning. Our findings indicate that probabilistic learning—the ability to extract predictable patterns from a visual stream—was not only immune to the negative effects of mind wandering but even appeared to benefit from such periods of inattention. Furthermore, improved probabilistic learning was linked to the nature of mind wandering: spontaneous, as opposed to deliberate mind wandering, was associated with better probabilistic learning further supporting the link between transitory, uncontrolled lapses of attention and model-free computations. Notably, the associations between mind wandering and probabilistic learning performance were present at the intraindividual level: when participants experienced mind wandering relatively more and more spontaneously (as referenced to their own average), they appeared to exhibit better probabilistic performance. Such associations were not observed at the interindividual level; therefore, our results do not support the notion that dispositional differences in mind wandering were predictive of increased probabilistic learning. Finally, oscillatory activity in the low (slow and delta) frequency range indicative of covert sleep-like states was associated with both mind wandering and probabilistic learning in the first half of the task. Considering the key role of probabilistic learning in shaping behavior and underlying neural computations (Santolin and Saffran, 2018; Fiser and Lengyel, 2019), our findings add further insights into the benefits of task-unrelated thoughts in human cognition.
The benefit of mind wandering in probabilistic learning may seem surprising in light of the large number of studies highlighting the costs of task-unrelated thoughts. Nevertheless, our finding does not stand alone without empirical antecedents. Emerging studies indicate that loosening attention and engaging in mind wandering or even entering into the transition between wakefulness and sleep onset might facilitate the processing of hidden regularities in the perceptual information stream which are less accessible for model-based, goal-directed attention (Thompson-Schill et al., 2009; Amer et al., 2016; Decker et al., 2023; Vékony et al., 2025). Enhanced probabilistic learning in the ASRT task was previously shown to be associated with periods of mind wandering in an online study (Vékony et al., 2025). The present results, replicating, extending, and validating this finding in a laboratory environment, point to the robustness of this link. Of note, the implementation of the ASRT task in the current study was different from that of Vékony et al. (2025) in terms of stimulus type/spatial location, as well as pacing, suggesting that the beneficial effect of mind wandering on probabilistic learning generalizes across different contexts and task versions. Our findings accentuating a link between mind wandering and probabilistic learning are also in line with studies that highlighted the role of the default mode network (DMN) in implicit learning and the extraction of statistical regularities. More specifically, sustained and coordinated activity in key regions of the DMN (a large-scale network associated with task-unrelated, self-referent thoughts; Buckner and Carroll, 2007; D’Argembeau et al., 2007; Raichle and Snyder, 2007; Smallwood and Schooler, 2015) was shown to contribute to performance in different implicit learning tasks (Yang et al., 2010; Kóbor et al., 2024) as well as to be involved in automated responses after the acquisition of simple statistical rules (Vatansever et al., 2017).
Better probabilistic learning under periods of mind wandering is in line with the assumptions of the competition framework suggesting an antagonistic relationship between model-based and model-free mental processes or more specifically, between cognitive control and probabilistic learning (Poldrack and Packard, 2003; Janacsek et al., 2012; Pedraza et al., 2024). The competition framework postulates that probabilistic learning is enhanced under conditions of reduced executive control, an assumption that has received ample empirical support in recent years (Nemeth et al., 2013b; Virag et al., 2015; Ambrus et al., 2020; Pedraza et al., 2024). Since task-unrelated thoughts are consistently linked to reduced executive control (Smallwood et al., 2008; McVay et al., 2009; McVay and Kane, 2010, 2012a; Kam and Handy, 2014; Kawagoe, 2022), we may speculate that improved probabilistic learning under periods of mind wandering might reflect an automatic shift to model-free from model-based forms of learning. By broadening the scope of attention, mind wandering might facilitate effortless and unconstrained exploration of the environment and promote favorable behaviors in situations where task rules and goals are somewhat opaque. The uncontrolled occurrence of task-unrelated thoughts might reflect such a shift in cognitive processes and hence, facilitate the extraction of hidden (probabilistic) dependencies between stimuli.
Notably, mind wandering was linked to better probabilistic learning when it was experienced as spontaneous. Spontaneous task-unrelated thoughts seem to be different from deliberate mind wandering: the latter is characterized by the intention to engage in (usually) pleasant thoughts, feelings, and fantasies in a more constrained, goal-directed manner, while the former reflects the unintentional drifting of attention from the current task (Seli et al., 2015). Although the mechanisms differentiating spontaneous from deliberate mind wandering are not fully understood, spontaneous mind wandering is more consistently associated with executive failures and inattention, whereas engaging in deliberate mind wandering appears to be linked to reduced motivation to maintain task focus (Bozhilova et al., 2018; Robison and Unsworth, 2018; Robison et al., 2020). Interestingly, when compared with mind wandering with meta-awareness, episodes of task disengagement that remained unnoticed were associated with increased activity in several regions of the frontoparietal control network (Christoff et al., 2009), possibly providing an explanation for the stronger interference of spontaneously evolving mind wandering with neural mechanisms of task-related cognitive control. Accordingly, our findings revealed distinct associations of the type of mind wandering with task performance. Whereas spontaneous mind wandering was linked to better probabilistic learning, deliberate mind wandering was associated with enhanced visuomotor accuracy. We may speculate that deliberate mind wandering was less detrimental to task performance as the task did not demand further attentional resources.
Our findings observing increased oscillatory activity in the slow and delta frequency ranges linked to mind wandering aligns with the study of Andrillon et al. (2021) reporting an association between slow waves, mind wandering, and impaired task performance. Increases in slow frequency activity in connection to mind wandering are not a unanimous finding in previous studies (Kam et al., 2022). Nonetheless, Andrillon and colleagues, instead of focusing on conventional band-wise power analyses, detected individual slow waves, providing a more sensitive method to capture transient EEG oscillations. Our approach to distinguish aperiodic and periodic EEG components appeared to be sensitive in capturing the slow frequency correlates of mind wandering in contrast to traditional (band-wise) EEG analyses, which, due to the intermingling of aperiodic and periodic components, might mostly be able to unravel more robust oscillatory activities, particularly in the alpha range.
Mind wandering associated with inattention and deficits in executive control is thought to be triggered (at least partly) by transient, sleep-like activity expressed in increased slow frequency oscillations (Andrillon et al., 2019; Jubera-Garcia et al., 2021; Wienke et al., 2021). Our findings not only provide further evidence for this claim but also suggest that these so-called offline states may have consequences beyond the negative impacts on cognition and behavior. Offline states, accompanied by sleep-like cortical activity and experienced as mind wandering might facilitate the rapid consolidation of previously acquired information, similar to sleep-related memory consolidation. A large body of evidence has been accumulated during the last three decades regarding the beneficial influence of postlearning sleep on information processing (Diekelmann and Born, 2010; Rasch and Born, 2013). Sleep seems to be an ideal state for memory consolidation because of (1) the attenuation of environmental stimuli and online encoding processes, (2) network activity at different neural levels (hippocampal, thalamocortical, corticocortical) facilitating the reactivation and reconsolidation of memories (Rasch and Born, 2013; Stickgold, 2013), and (3) optimization of signal-to-noise ratio by the readjustment of synaptic weights (Tononi and Cirelli, 2006). Although the idea that memories are specifically strengthened during sleep is intellectually appealing, recent studies question the exclusive role of sleep in consolidating previously encoded material and suggest that periods of waking rest are equally beneficial for memory consolidation (Pan and Rickard, 2015; Wamsley, 2019, 2022; Németh et al., 2024; Ogawa et al., 2024). Accordingly, key regions of the DMN exhibiting coherent activity during off-task, resting states (Raichle, 2015) seem to be involved in memory processing (Kaefer et al., 2022) and were associated with different neural markers of memory reactivations in both animals (Kaplan et al., 2016) and humans (Higgins et al., 2021).
It is tempting to speculate that enhanced probabilistic learning associated with mind wandering may be linked to local sleep boosting offline memory consolidation in the waking brain (Wamsley and Summer, 2020; Wamsley, 2022). This assumption extends classical theories of memory consolidation, such as sleep-dependent and time-dependent memory consolidation, proposing a new type of consolidation termed as local sleep-dependent consolidation (Vékony et al., 2025). Our assumption aligns with the opportunistic memory consolidation theory, which suggests that consolidation of memories can occur while awake, asleep (Mednick et al., 2011), or—as we speculate—under local sleep states. Furthermore, local sleep-dependent consolidation may not only stabilize previously encoded material but also support the formation of predictive representations. This way, such transient offline periods may facilitate the extraction of regularities and underlying probabilities, which are essential to predictive processes. However, further studies employing magnetoencephalography or intracerebral recordings are needed to test this idea, providing direct evidence of the relationship between mind wandering, local sleep, and the formation and updating of predictive representations.
In sum, we hypothesize that model-free learning, such as implicit probabilistic learning, might not even require periods of waking rest but could undergo rapid consolidation during task acquisition, especially when participants enter into a state of mind wandering (i.e., a transient offline state). The positive association between probabilistic learning and mind wandering during task acquisition may also explain why probabilistic learning occurs mainly through practice (Quentin et al., 2021), especially if postpractice rest periods are limited (Szücs-Bencze et al., 2023), and why probabilistic learning does not seem to benefit further from postlearning sleep or long periods of rest (Nemeth et al., 2010; Simor et al., 2019). Notably, in our study, the association between mind wandering and probabilistic learning showed an interaction with the time spent engaging with the task. The positive link between mind wandering and probabilistic learning was observed in the beginning, whereas it could not be detected at the second part of the task. Such pattern is in line with the assumption that experiencing mind wandering during task acquisition may be paralleled by the extraction and rapid consolidation of probabilistic information hidden in the visual stream.
In line with the above, EEG correlates of mind wandering and probabilistic learning showed considerable topographical overlaps in the first half of the task. In particular, increases in slow and delta oscillatory activity, most prominently in centroparietal locations (pointing to the involvement of sensorimotor regions; Melnik et al., 2017) were linked to enhanced mind wandering and probabilistic learning.
Since low-frequency activity during sleep and waking rest seem to have key roles in offline memory consolidation, our observation points to the potential influence of covert sleep states on mind wandering and probabilistic learning. These findings are interesting because in the ASRT task, the raw probabilities and statistics are extracted from the visual information stream precisely in the first half of the task (Janacsek and Nemeth, 2012). Hence, slow and delta oscillatory activity, along with mind wandering, may support the discovery of probabilistic patterns within the information stream. Nevertheless, whether centroparietal low-frequency oscillations reflect covert and local sleep-like states over sensorimotor regions and play a role in facilitating information processing remains to be explored. Similarly, whether oscillatory activity associated with mind wandering is linked to activations in the default mode network (DMN) and its role in memory reactivations warrants further investigation. Notably, although probabilistic learning continued to improve with extended practice (i.e., in the second part of the task), mind wandering no longer seemed to benefit task performance. We speculate that mind wandering may have helped participants extract the underlying statistics early in the task, while further practice contributed to the stabilization of probabilistic information. However, future studies are needed to investigate the mechanisms and distinct neural correlates of these potentially different phases of probabilistic learning. In addition, further studies should also systematically manipulate oscillatory activity and/or sleep pressure during task performance to infer the mechanistic role of oscillatory activity in mind wandering and information processing.
The EEG correlates of mind wandering, however, were not limited to low-frequency oscillatory activity. In the first half of the task, a widespread increase in beta oscillatory activity was observed. In addition, in the second half of the task, mind wandering was associated with the aperiodic component, more specifically with steeper spectral slopes in posterior regions, as well as with increased centroparietal theta and alpha, and an increase in the beta band in midline locations along the anteroposterior axis. Enhanced spectral power in the theta and alpha ranges were observed in previous studies as correlates of mind wandering (Kam et al., 2022). Interestingly, the EEG correlates of mind wandering in the second half of the task resemble the pattern observed in relation to the influence of time spent practicing the task (i.e., the effect of block number). A steeper spectral slope reflecting enhanced inhibition over excitation as the background activity and increased periodic activity in theta-alpha range might reflect perceptual decoupling and a drift in attention toward self-generated thoughts. Although studies examining the relationship between aperiodic and periodic components and cognitive performance remain limited, our findings align with a recent study showing a flattening of spectral slopes during tasks requiring selective attention (Waschke et al., 2021). Additionally, oscillatory activity in the alpha range was associated with reduced cortical excitability (Romei et al., 2008), presumably linked to inhibition and reallocation of cognitive resources (Jensen and Mazaheri, 2010). Shifting attention to the internal world from the external environment (and engaging in mind wandering) might be more prevalent when visuomotor control becomes more and more automatic as the task progresses. Oscillatory activity in the beta range was positively associated with mind wandering in both the first and the second half of the task. Interestingly, beta power was not consistently related to mind wandering in previous studies (exhibiting increases, decreases or, null effects; Rodriguez-Larios et al., 2021; Kam et al., 2022; Musat et al., 2024). Nevertheless, increases in beta power were associated with the time spent with a monotonous task (Perrier et al., 2016), resembling our results showing increased beta activity as the ASRT task progressed. Moreover, performing vigilance tasks after sleep deprivation led to reduced performance and concomitant increases in beta power (Lorenzo et al., 1995; Corsi-Cabrera et al., 1996; Mairesse et al., 2009) that was assumed to reflect cognitive efforts to compensate for accumulating fatigue, microsleep, and proneness to cognitive errors (Mairesse et al., 2009). Further research is needed to determine whether our findings, showing an association between mind wandering and beta activity, reflect such compensatory efforts. We should note, however, that it may seem difficult to compare the EEG correlates of mind wandering in tasks which differ in cognitive load, cognitive processes, or focus of attention. That is, the phenomenological aspects, the neural correlates, as well as the functions of mind wandering might considerably differ in different contexts. Moreover, previous studies examining the EEG spectra during mind wandering mostly focused on band-wise spectral power overlooking and intermingling the aperiodic and periodic components of the EEG spectrum (Donoghue et al., 2020). Whereas the aperiodic component reflects background activity and neural excitability (i.e., excitation/inhibition balance; Gao et al., 2017), periodic components may capture moment-to-moment changes in neural firing in response to external stimuli or internally driven cognitive processes (Csercsa et al., 2010; Halgren et al., 2019; Ujma et al., 2022).
Despite its potential benefits, mind wandering was also associated with cognitive costs: loosening task focus was linked to poorer visuomotor performance, that is, participants in general (regardless of the probabilistic nature of the trials) committed more errors in blocks when they reported to mind wander. Such findings are in line with previous reports accentuating the negative aspects of mind wandering (Mooneyham and Schooler, 2013; Smallwood and Schooler, 2015) and suggesting that task-unrelated thoughts disrupt task performance (Andrillon et al., 2019, 2021). Whereas the costs of mind wandering in attention-demanding tasks are widely established, future studies combining attentional or executive control with probabilistic learning in a single task could examine whether the costs of mind wandering in attention and control also convey its benefits in other domains, such as probabilistic learning. Our study is not without limitations. Since we assessed mind wandering only at the end of each block, our database contained a relatively limited number of self-reports. While increasing the number of thought probes throughout the task would have been advantageous, frequent sampling during practice could heighten metacognition and interrupt the stream of perceptual stimuli, potentially interfering with task performance (He and Li, 2023). Unlike in other cognitive tasks (e.g., SART), thought probes in our study were not randomly distributed but always appeared at the end of blocks, which could have introduced expectancy effects. However, participants were unaware of the exact block duration and could only subjectively estimate when each block would end. This makes it unlikely that they could precisely anticipate the occurrence of thought probes. Although randomizing thought probe timing could have further reduced expectancy effects, it might have disrupted the learning process occurring during the uninterrupted 1.5–2 min practice blocks (Quentin et al., 2021). Importantly, our task differs from others that follow a repetitive, fixed sequence of stimuli; in the ASRT task, the sequence follows a probabilistic pattern. Interrupting these practice blocks with thought probes might have heightened participants’ metacognitive awareness of mind wandering, which could interfere with the automatic acquisition of probabilistic information. Moreover, although we explored the types of task-unrelated thoughts to some extent, the mental experience of mind wandering is evidently more complex and presumably way more heterogeneous. Future studies should assess the phenomenological aspects of mind wandering in more detail in order to disentangle the potential benefits and costs of task-unrelated thoughts on task performance. For instance, individuals seem to show pronounced differences in their tendency and nature of mind wandering. Some experience more often positive-constructive daydreaming while others are more prone to negative, guilty-dysphoric styles (Blouin-Hudon and Zelenski, 2016). Studying the content and affective nature of mind wandering may provide a more nuanced view about the links between task-unrelated thoughts, negative affect, and cognitive performance (Killingsworth and Gilbert, 2010; Mooneyham and Schooler, 2013). Moreover, spontaneous thoughts might have different temporal direction (e.g., past vs future-oriented thoughts), which may also have a distinct association with task performance. While several studies examined the phenomenological aspects of mind wandering such as the vividness, bizarreness, or narrative structure of daytime mental contents (Gross et al., 2021), whether such dimensions have a particular association with cognitive performance and cortical activity is still unknown. In addition, our sample had an uneven distribution of females and males and consisted of young university students, limiting the generalizability of our findings. Although gender differences do not appear to play a significant role in probabilistic learning (Juhász and Nemeth, 2018) and mind wandering (Mowlem et al., 2019), the underlying neural correlates may vary (Bertossi et al., 2017). Therefore, future studies should aim for a balanced gender ratio and include participants from more diverse backgrounds. Despite these limitations, our study provides empirical support for the potential benefits of mind wandering in information processing and indicates that although visuomotor accuracy is reduced, probabilistic learning is enhanced during periods of attenuated task focus. Our study suggests that mind wandering and transient, local sleep-like states might facilitate the rapid extraction and consolidation of hidden regularities from the visual information stream. This phenomenon may provide the basis for the acquisition of new skills and optimization of predictive processes.
Footnotes
This work was supported by the Chaire de Professeur Junior Program by INSERM and French National Grant Agency (ANR-22-CPJ1-0042-01); the National Brain Research Program project NAP2022-I-2/2022 (D.N.); Hungarian National Research, Development and Innovation Office Grant NKFI FK 142945 (P.S.); and Janos Bolyai scholarship of the Hungarian Academy of Sciences (P.S.).
The authors declare no competing financial interests.
- Correspondence should be addressed to Teodóra Vékony at teodora.vekony{at}pdi.atlanticomedio.es or Péter Simor at simor.peter{at}ppk.elte.hu.
