Ultraslow brain dynamics as neurophysiological markers of information sampling during learning Authors: Patricia León-Cabrera, Anna Prunell-Castañé, Thomas F Münte, Shunkai Wang, Antoni Rodríguez-Fornells Published: 2026 ABSTRACT SummaryTo learn, we must actively sample information from our environment: selecting, prioritizing, and extracting relevant cues. Here, we investigated whether ultraslow event-related potentials (ERPs) could track information sampling as participants learned new words from sentence contexts that provided varying amounts of information about their meaning, requiring them to actively extract learning-relevant cues. Ultraslow negative ERPs emerged while participants engaged in learning. Although ultraslow brain activity was not modulated by the level of informativeness of the contexts, individual differences in learning success were associated with steeper ultraslow voltage slopes. Specifically, participants with better learning outcomes exhibited a steeper ultraslow slope while learning from informative contexts, suggesting greater engagement and exploitation of relevant information. Together, these findings identify ultraslow brain activity as candidate neurophysiological markers of information sampling and provide evidence that individual differences in these dynamics relate to successful learning.Keywords: information sampling, contextual word learning, ultraslow event-related potentials, N400 Introduction Learning involves actively sampling and exploring new information, with learners continuously evaluating which pieces are most relevant to their learning goals. This prioritization process becomes especially critical in the absence of explicit instruction, where individuals must independently gather and extract meaningful content to succeed, and relies on the interplay between attention, memory, and reward systems.1,2 Despite the central role of effective information sampling in self-regulated learning, its neural correlates remain poorly understood, and to date, no research has tracked their temporal dynamics. An ecological situation to study information sampling is contextual word learning: the acquisition of the meaning of a new word from context. Nagy and Gentner3 liken contextual word learning to dipping a magnet into a mixture of iron filings and sand—where the iron filings (relevant information) stick, and the sand (irrelevant information) falls away. In line with this analogy, previous studies employing electroencephalography (EEG) have demonstrated that learners can effectively and rapidly identify relevant information in the context to derive the meaning of a novel word.4,5 Many of these studies have capitalized on the N400 event-related potential (ERP), a negative-going voltage deflection appearing 300–500 ms after the onset of a meaningful stimulus,6 as it is associated with lexical-semantic processing during comprehension.7 For example, Mestres-Missé et al.4 tracked changes in brain activity elicited by novel words (Spanish pseudowords) as learners discovered their meaning from contexts, and found that, by the third exposure of a novel word, it elicited an N400 component that was comparable to that of real words, which was taken as a marker of successful meaning acquisition.8,9,10 Importantly, the amplitude of the N400 is positively correlated with word predictability11,12,13 and has been successfully modeled as reflecting semantic prediction error—the discrepancy between predicted and actual semantic content.14 In line with this, several contextual world learning studies have shown reduced N400 amplitudes to novel words presented in meaningful contexts (relative to meaningless or low constraining ones),15,16,17 indicating smaller prediction errors when the context supports meaning acquisition. Also, N400 amplitudes are reduced for words primed by successfully learned novel words.4,5 Taken together, these findings suggest that learners effectively sample information during contextual word learning without explicit instruction. While the aforementioned N400 studies provide valuable insights, they do not show how the process of sampling information operates continuously in time, as learners are processing the sentences. Indeed, sentence comprehension involves continuously integrating information to build a high-level representation (internal model) of the meaning of the sentence.18,19,20 As information is gathered and certainty about the meaning increases, individuals can use this internal model to exert top-down influence on processing by pre-activating information and predictively allocating an adequate amount of resources.21,22,23 For example, evidence shows that strongly informative contexts are associated with the generation of knowledge-based inferences24,25,26,27 and the pre-activation of incoming words.28 In line with this, a contextual word learning task employing eye-tracking showed that, after being presented with a novel word, learners spent more time reading incoming context segments that were informative (“Joe picked up the asthor and began to strum a tune”) than those that were uninformative (“Joe picked up the asthor and began to walk home”) about its meaning.29 This extra processing time in informative contexts was attributed to enhanced cognitive effort to extract and integrate meaning and generate elaborative inferences. Furthermore, after informative (vs. uninformative) contexts, learners spent less time reading the superordinate term of the novel word (“He played the instrument to relax”), a result that is in line with prediction (for a review, see Kutas et al.30). A more recent eye-tracking study found that working memory (WM) has a central role in contextual word learning, as individuals with higher WM capacity were more effective at retaining relevant contextual cues and integrating them with the novel word, resulting in better learning outcomes.31 Therefore, prior evidence indicates that successful meaning acquisition during contextual word learning is associated with greater online investment of cognitive resources. Noteworthily, at the neurophysiological level, online changes in cognitive demands during sentence processing have been tracked through slow and ultraslow ERP modulations.32 Several studies have consistently reported slow negative-going ERPs during word-by-word sentence processing associated with enhanced cognitive processing,33 particularly increased WM demands.33,34,35,36,37,38 Earlier work by King and Kutas39 initially proposed that these ultraslow components might reflect different sentence-level processes, as for example (i) continuous encoding of rapidly presented visual information, (ii) the built-up of high-level sentence-level representations from the encoded word-level input, or (iii) the temporary storage of currently unintegrated material within WM. More recently, León-Cabrera et al.40 showed that slow (below 5 Hz) negativities emerged progressively over the course of sentence contexts that allowed the prediction of an incoming word (“The goalkeeper managed to catch the … ball”) and interpreted it as reflecting increased WM demands linked to top-down prediction. This sentence-level ultraslow negativity might belong to a broader family of prediction-related negativities with variable durations that precede words when informative contexts or cues allow their prediction41,42,43,44 and that have been associated with the engagement of the beforementioned predictive operations, including the pre-activation of upcoming linguistic input and increased anticipatory attention (for a review, see León-Cabrera et al.45). Based on this evidence, the present study aimed to characterize the temporal dynamics of neurophysiological activity underlying information sampling during learning by examining ultraslow ERP modulations. To investigate this, we designed a contextual word learning task in which participants had to infer the meaning of a novel word (a Spanish pseudoword) embedded in a three-sentence context (see Figure 1). The first sentence introduced the novel word in its final position (“He likes to walk in the ánaro”). The second sentence was either informative (“He enjoys it when the weather is hottest.”) or uninformative (“He has always been great with numbers.”) about the word’s meaning (i.e., summertime). Finally, the third sentence presented the superordinate term (“season”) of the novel word at a predictable position. As a control, we included a condition where the first sentence contained a real word (“summertime”) instead of the novel word and therefore did not involve learning. Materials and design (A) Examples of three-sentence contexts for each condition. Horizontal lines below sentences indicate the time-locking event to investigate N400 effects (blue lines) and ultraslow modulations (black line). Please note that these are approximate English translations that fit the number of words per sentence. (B) Depiction of the structure of a task block. On the left, each box in represents the presentation of a word (e.g., first word is w1, second word is w2, etc.) in chronological order. Each row of boxes corresponds to a sentence. Boxes with blue contours indicate words for which N400 effects were computed. On the left, a depiction of the forced-choice test that participants completed at the end of each block. After selecting one of the four options, the selected word changed to green if it was correct (top screen) or to red if it was incorrect (bottom screen). We examined the development of ultraslow ERP modulations (below 0.7 Hz) while learners read the second and third sentences word-by-word. We hypothesized that ultraslow negative potentials would emerge in this period prior to the superordinate term, with greater amplitudes in the informative condition (Info) compared to the uninformative condition (Uninfo) and the control condition (Control), reflecting resource allocation driven by effective information sampling. In addition, we examined N400 responses at the superordinate term, predicting reduced amplitudes in the informative condition, indicative of lower semantic prediction error. Finally, we investigated individual differences in information sampling strategies by correlating these neurophysiological markers with learning outcomes, measured during the task through a forced-choice test at the end of every experimental block (Figure 1). Results Ultraslow modulations Figure 2 shows the grand-average ERPs to the processing of the second and third sentences of the sentence triplets, presented word by word. Each word elicited an N100 component, followed by a P200 component and an N400 component. As expected, the application of a low pass filter (below 0.7 Hz) to focus on ultraslow modulations revealed an ultraslow negativity that unfolded gradually over the course of the second sentence and plateaued before the presentation of the superordinate word (i.e., fourth word of the third sentence) in all conditions. Importantly, the same pattern was observed when a less aggressive low-pass filter (<5 Hz) was applied (see Figure 2B), suggesting that the ultraslow modulation does not reflect a filtering artifact. Grand average ERPs of the second and third sentences (A) Grand average ERP of all conditions at Cz electrode in the interval encompassing the second and third sentence. The plotted data are low-pass filtered at 5 Hz. The seventh word (W7) was the last word of the second sentence, which was followed by a short break before the presentation of the third sentence. The eleventh word (W11) was the superordinate term of the word that had to be learned. (B) Grand average ERPs of all conditions applying different low-pass filters at several electrodes (F3, Fz, F3, C3, Cz, C4, P3, Pz, and P4) to demonstrate the result of applying a strong low-pass filter (<0.7 Hz) to capture ultraslow modulations, taking the informative condition as an example. Cluster-based permutation (CBP) tests (including all time points and electrodes) revealed no statistically significant differences between the info and uninfo conditions (all cluster p values > 0.46) and therefore no effect of the informativeness manipulation on ERP amplitudes in the interval encompassing the second and third sentences (Figure 3). Instead, amplitudes were smaller in the control relative to the info condition (p = .015; cluster mass = −44.068) and to the uninfo condition (p = .013; cluster mass = −49.557), suggesting a general effect of learning on ultraslow activity regardless of the level of informativeness of the contexts. The same pattern of results emerged when a linear mixed effects (LMEs) model was fitted on the data (see Document S1). Cluster-based permutation results on ultraslow activity Raster plots showing the results of the cluster-based permutation analyses (including all timepoints and electrodes). Vertical gray lines indicate the onset of every word, which are labeled on top (e.g., W1 = first word). Blue segments indicate statistically significant t values (range: −2 to −5; p < 0.05) at every time-point and electrode within the cluster with the maximum cluster-level statistic between info and control contexts and uninfo and control contexts. Below each raster plot, scalp maps of the mean voltage amplitude difference between conditions at every word (500 ms long interval) are shown. Critically, each contrast showed a different time-course. In the control vs. info contrast, the conditions differed statistically already at the first word (sentence 2, W1) and lasted until the onset of the superordinate word (sentence 3, W4), whereas in the control vs. uninfo contrast, the cluster started later, at the third word (sentence 2, W3), and differences lasted until the end of the sentence triplet (sentence 2, W7). Next, we investigated the relation between ultraslow modulations and learning performance. As shown in Figure 4A, the mean percentage of correct responses in the forced-choice test was 56.3% in the info condition and 46.5% in the uninfo condition at the group-level. Therefore, performance was above chance in both, but better for new words in the info compared to the uninfo condition, t(44) = 5.71, p <. 001; 95% confidence interval = 6.3%–13.3%; mean of the differences = 9.81). In addition, participants who performed better in the info condition also performed better in the uninfo condition as indicated by a statistically significant positive correlation between the percentage of correct responses in both conditions, r(45) = 0.633, p < 0.001. Relation between ultraslow modulations and learning performance (A) Violin plots comparing the distribution of the mean percentage of correct responses in the forced-choice test as a function of the condition in which the target item was presented (info or uninfo). Performance differed between conditions (p < .001). Horizontal lines inside the violin plots correspond to the median (continuous line) and interquartile range (discontinuous lines). Wider sections of the violin indicate a higher probability of observations taking that value than thinner sections. The discontinuous horizontal line indicates the chance level (25%). (B) Ultraslow modulation trajectory as a function of overall learning performance. Estimated mean centroparietal amplitude (in microvolts) across word positions, derived from the LME model. Lines represent model-estimated trajectories at three representative levels of overall learning performance: worse learners (−1 SD; red), average learners (0 SD; gray), and better learners (+1 SD; green). Shaded areas indicate 95% confidence intervals. (C) The horizontal double-sided arrow indicates the interval in which the slope was computed for every participant and condition. (D and E) Scatterplots of the significant correlation (alpha level = .05) between learning performance (y axis: percentage of correct responses in the forced-choice test) and slope of the ultraslow negativity (x axis; in microvolts) in the info condition and in the uninfo condition. Each data point represents a participant. We correlated the slope of the ultraslow negativity in the second sentence (from the onset of the second sentence to the penultimate word) with the percentage of correct responses separately for info and uninfo conditions (Figure 4C). There were no differences in slope between the info (mean slope = –0.794, SD = 0.882) and uninfo condition (mean slope = –0.878, SD = 0.952), t(44) = 0.853, p = 0.398, at the group-level. This was expected given the results of the cluster permutation tests, which revealed that the two conditions did not differ statistically in this interval. However, correlational analyses revealed a statistically significant correlation (alpha level = 0.05) between the slope of the ultraslow negativity and learning success only in the info condition. Specifically, steeper negative slope when reading informative contexts was associated with a higher percentage of correct responses in that same condition, r(43) = –0.37, puncorrected = 0.014, pHB = 0.027 (Figure 4D). This correlation fell slightly below the threshold indicated by our sensitivity analysis as necessary for 80% (N = 45; α-level = .05, two-tailed); importantly, however, the 95% percentile bootstrap CI did not include zero, [–0.58, –0.14], indicating that the association is unlikely to be null. The same trend was observed in the uninfo condition, but the correlation did not reach statistical significance, r(43) = –0.190, puncorrected = .221, pHB = 0.205, 95% percentile bootstrap CI [–0.47, 0.12] (Figure 4E). To complement the correlational analyses, we ran an LME model on the word-by-word trajectory of ultraslow modulations, allowing us to capture the signal at a finer temporal resolution and to test whether the relation generalized to overall learning performance in the task. The model revealed a significant main effect of time, F(1, 446) = 238.17, p < 0.001, reflecting a robust negative-going drift in centroparietal amplitude across the second sentence interval. A significant main effect of overall learning was also found, F(1, 43) = 7.51, p = 0.009, indicating that better learners showed overall more negative centroparietal amplitude. The time × learning interaction was significant, F(1, 446) = 28.42, p < 0.001, indicating that the rate of the negative-going drift increased with learning performance. Specifically, the linear slope was significantly negative at all performance levels and steepened progressively with performance. To probe this interaction, we estimated the linear slope at three representative values of learning performance (Figure 4B): worse learners (−1 SD): β = −1.10, t(446) = −7.06, p < 0.001; average learners (0 SD): β = −1.70, t(446) = −15.43, p < 0.001; better learners (+1 SD): β = −2.29, t(446) = −14.64, p < 0.001. No other effects reached significance. N400 component As an important control, N400 amplitudes (300–500 ms post-word onset) to the new words (i.e., last word of the first sentence) were comparable in the info and uninfo conditions (Figures 5A and 5C). Also, replicating previous findings, real words (i.e., target item in the control condition) elicited larger N400 responses than new words at first exposure,4 both in the info condition (p = 0.022; cluster mass = −30.62) and in the uninfo condition (p = 0.002; cluster mass = −46.44) (Figure 5E). N400 effects (A and B) Grand average ERPs of each time-locked (0–800 ms) to the target item and to the superordinate word at a set of representative electrodes. The grey-coloured areas indicate the interval (300–500 ms) where cluster-based permutation analyses were run (alpha level = 0.05; two-tailed testing). Data were low-pass filtered at 12 Hz for plotting purposes. (C and D) Scalp maps showing mean amplitude differences in the 300 to 500 ms interval for statistically significant contrasts at the target item (control minus info; control minus uninfo) and at the superordinate word (uninfo minus info; uninfo minus control). Highlighted electrodes (white dots) indicate statistically significant differences in the cluster-based permutation analysis (p < 0.05). (E) Bar plots displaying mean amplitudes (in microvolts) in each condition at the Pz electrode in the 300–500 ms after the target item and (F) after the superordinate word. Data are represented as mean ±1 SD. Critically, the presentation of the superordinate word (fourth word of the third sentence) elicited larger mean amplitudes in the N400 interval in the uninfo than in the info condition (Figures 5B and 5D). This effect of informativeness on the N400 amplitude was captured by a significant negative cluster that was most prominent at centro-parietal sites (p < 0.001; cluster mass = −77.11) (Figure 5D). Likewise, although marginally significant, mean amplitudes were larger in the uninfo condition than in the control condition, as reflected by a right posterior cluster (p = 0.049; cluster mass = −20.12) (Figure 5D). There were no detectable differences between the info and the control condition in the analyzed interval (all cluster p values above 0.14) (Figure 5F). Finally, there were no statistically significant correlations between the N400 informativeness effect at the superordinate word (mean amplitude differences 300–500 ms post-word onset) and learning performance in the info condition, r(43) = 0.052, puncorrected = 0.734, pHB >1, as well as the uninfo condition, r(43) = −0.04, puncorrected = 0.763, pHB >1. There were also no correlations between learning performance and the N400 amplitude difference between the control condition and the info condition, r(43) = 0.298, puncorrected = 0.046, pHB = 0.223, or the control condition and the uninfo condition, r(43) = 0.216, puncorrected = 0.153, pHB = 0.614. Discussion This study investigated the temporal dynamics of neurophysiological activity underlying information sampling by tracking ultraslow ERPs during a word learning task. To address this, we collected participants’ brain responses while they learned the meaning of new words from written contexts. Sentence contexts conveyed different amounts of information (informative or uninformative) about the new word’s (“ánaro”) meaning (“summertime”). Crucially, an ultraslow negativity emerged gradually during sentence contexts in both learning conditions (relative to a control, no-learning condition) before the presentation of the superordinate term (“season”), which provided critical information for meaning acquisition. Although the amplitude of the ultraslow ERPs did not differ as a function of contextual informativeness, its steepness scaled with learning outcomes: participants with better overall learning in the task showed a greater word-by-word increase in ultraslow modulation amplitude, and those with better learning outcomes in the informative condition exhibited a steeper slope of the ultraslow negativity in these contexts, suggesting that they engaged with and exploited informative contexts to a greater extent than less successful individuals in that condition. Notably, participants were rather successful in discovering the meaning of new words from the given contexts in learning conditions (Info and Uninfo). On average, performance was above chance in both conditions of contextual word learning. However, as expected, successful recognition was higher for words in informative contexts.4,15 Even though informative contexts were more constraining about the superordinate term (e.g., “flower”) than about the specific member of the category (e.g., “rose”), participants could discern the correct meaning in the forced-choice test, with some individuals exceeding 75% of accuracy in the informative condition. This behavioral advantage for informative over uninformative contexts was mirrored at the neurophysiological level: the N400 response to the superordinate word was significantly reduced in informative contexts compared to uninformative ones, indicating facilitated processing due to prediction.7,14 Moreover, the lack of N400 differences between informative and control contexts suggests that informative contexts effectively brought prediction error reduction down to levels typical of highly predictable lexical input as a result of successful meaning acquisition. In relation to the main goal of the study, we examined ultraslow ERP activity (below 0.7 Hz) during sentence processing and found that it distinguished between learning and non-learning situations. In both learning situations (informative and uninformative conditions), a sustained slow negative buildup emerged prior to the presentation of a key piece of information: the superordinate term of the word those participants had to learn. The inclusion of the superordinate term in the contexts was unbeknownst to participants; however, since it consistently appeared at a fixed position (the fourth word of the third sentence), participants could anticipate its presentation if they recognized it as informationally relevant. In this view, the ultraslow negative ERP modulations found in the learning conditions (relative to the control condition) might reflect a heightened state of anticipatory attention toward the temporarily predictable and informationally relevant superordinate term. In other cognitive domains, non-motor anticipatory potentials like the stimulus-preceding negativity (SPN) are reliably observed before the presentation of an attended feedback stimulus that will provide relevant and motivationally salient information46,47 and, interestingly, its amplitude might index the amount of information that one expects to extract from the awaited stimulus.48,49 This interpretation is also in agreement with the proposal of Ferreira and Lowder21 that, as comprehenders achieve a good enough representation of sentence meaning, they can enter into a forward-looking state of “preparedness”—a heightened state of anticipatory attention toward new and non-redundant information, which might also involve a state of greater motivation. Interestingly, previous work has evidenced the impact of learning new words in reward-motivational systems.50,51 Overall, the emergence of these negativities indicates that learners in this task could successfully identify and sample key pieces of information during self-guided learning, in convergence with previous studies using eye-movements suggesting that learners are immediately sensitive to the availability of information about the meaning of novel words.29,31 The ultraslow ERP modulations found here show a centro-parietal distribution that contrasts with that of previous research on cross-sentence ERP modulations associated to WM, in which anterior left frontal34,35,37,39 or frontal anterior distributions have been observed.34,37,52 The topographical distribution, along with the morphological features—gradual overall increase in amplitude prior to the relevant stimulus and inversion of polarity afterwards—is reminiscent of semantic-related anticipatory negativities found in the domain of language comprehension prior to contextually predictable words.45 Specifically, the study by León-Cabrera et al.40 showed a similar centroparietal sustained negativity (<5 Hz) emerging across sentences that were strongly constraining about a sentence-final word, relative to weakly constraining sentences. In contrast, the current study did not reveal group-level differences in ultraslow activity between informative and uninformative contexts. Despite the null finding between informative and uninformative conditions at the group-level, a more nuanced pattern emerged when accounting for individual differences in learning outcomes. On one side, learning outcomes did not correlate with N400 effects, indicating that the minimization of semantic prediction error alone did not distinguish more from less successful learners in this task. By contrast, correlational analyses revealed that individuals who exhibited a steeper ultraslow modulation (i.e., greater slope) at the onset of the second sentence in informative contexts learned more from those contexts, and an LME leveraging the full word-by-word trajectory of the ultraslow modulation also revealed steeper ultraslow negativities in more successful learners. This is in line with the idea that learning success in this task relied on how actively and predictively participants sampled the incoming information21,23 which was better captured by the ultraslow activity modulations tracking the dynamics of information sampling continuously over time. Overall, this pattern of results reinforces the value of examining ultraslow dynamics alongside faster ERP components: unlike the N400, which indexes punctate operations time-locked to individual stimuli, ultraslow modulations unfold over seconds and are therefore uniquely positioned to track sustained and integrative processes, such as information sampling, as they develop in real time during learning. Altogether, these findings suggest that the amplitude and trajectory of the observed ultraslow negativities reflect a processing mode adopted by the learner: a steeper slope might reflect an individual’s ability to detect and engage with relevant information early on and devote greater cognitive resources to its processing.29 Yet, while the present data are compatible with this interpretation, ultraslow modulations likely encompass multiple cognitive processes involved in comprehension and could alternatively reflect a stable cognitive trait, such as attentional engagement or WM capacity.31 This is consistent with the broader literature on ultraslow ERP modulations, which remains limited and has yet to converge on a single functional account. Refining the mechanistic interpretation will likely require complementary measures and targeted experimental manipulations. Likewise, the current paradigm was deliberately designed to study information sampling under controlled experimental conditions, and more research is needed to determine if these results generalize to more naturalistic forms of self-guided learning. In conclusion, a gradual ultraslow negativity emerged in learning contexts but not in the no-learning control, indicating that learners successfully tracked relevant information during contextual word learning. While this signature was present during learning regardless of contextual informativeness, its steepness scaled with overall learning outcomes, and participants with better learning in informative contexts exhibited a steeper slope in those contexts specifically, suggesting that more successful learners engaged with and exploited informative contexts to a greater extent than less successful individuals. More broadly, these findings position ultraslow ERPs as candidate neurophysiological indices of information sampling during learning, offering a non-invasive window into how the brain dynamically allocates and sustains cognitive resources across extended timescales. Future work combining EEG with neuroimaging could help identify the cortical and subcortical generators of these dynamics and examine whether these dynamics generalize beyond controlled laboratory paradigms to more naturalistic learning situations, where information sampling unfolds over longer timescales and under richer, less constrained conditions. Ultimately, establishing reliable neurophysiological markers of information sampling could benefit research in domains that share a fundamental interest in understanding how individuals differ in their ability to extract and retain relevant information from their environment. Limitations of the study Some methodological limitations must be considered when interpreting these findings. One possibility is that the fixed position of the superordinate term encouraged participants to rely on temporal predictability rather than contextual content. However, our data argue against a purely structural-anticipation account: learning performance was significantly higher in the informative condition, and ultraslow negativities were larger in both learning conditions compared to the control condition despite identical temporal structures across all three. This indicates that the observed effects depend on the learner’s engagement with meaning extraction rather than positional expectancy alone. At the same time, the nature of this engagement may have been influenced by semantic coherence. To control for lexical and syntactic variables, we used identical second sentences across conditions, which entailed a trade-off: reduced semantic similarity across sentences in the uninformative condition relative to the informative and control conditions (for detailed results, see Figure S7). While this difference in semantic coherence may have influenced cognitive processing, it is unlikely to be the primary driver of the observed ultraslow ERP effects, as both learning conditions produced similar ERP patterns relative to the control, despite their internal disparity in semantic similarity. These considerations notwithstanding, the influence of these factors warrants careful attention in future research. Future studies should aim to disentangle them more explicitly, either by manipulating semantic coherence independently or by treating informativeness as a parametric variable rather than a binary classification. Furthermore, conducting planned analyses on the specific temporal intervals and spatial locations identified here would allow for a more granular assessment of how these distinct cognitive contributions unfold. Resource availability Lead contact Further information and requests for resources should be directed to the lead contact, Antoni Rodríguez-Fornells (antoni.rodriguez@icrea.cat). Materials availability This study did not generate new unique reagents. Data and code availability •The anonymized EEG data have been deposited at Open Science Framework (OSF) and are publicly available at https://doi.org/10.17605/OSF.IO/AVY4W as of the date of publication. DOIs are also listed in the key resources table. The anonymized EEG data have been deposited at Open Science Framework (OSF) and are publicly available at https://doi.org/10.17605/OSF.IO/AVY4W as of the date of publication. DOIs are also listed in the key resources table. •All original code used for EEG data preprocessing and analysis have been deposited at OSF and is publicly available at https://doi.org/10.17605/OSF.IO/AVY4W as of the date of publication. DOIs are also listed in the key resources table. All original code used for EEG data preprocessing and analysis have been deposited at OSF and is publicly available at https://doi.org/10.17605/OSF.IO/AVY4W as of the date of publication. DOIs are also listed in the key resources table. •Any additional information required to reanalyze the data in this article is available from the lead contact upon request. Any additional information required to reanalyze the data in this article is available from the lead contact upon request. Acknowledgments We would like to thank David Cucurell for his technical assistance during data acquisition. This study was funded by Ministerio de Ciencia e Innovación, through the Generación de Conocimiento grant, number PID2021-127130NB-I00, https://doi.org/10.13039/501100011033/ and ERDF - A way of making Europe. We thank CERCA Programme/Generalitat de Catalunya for institutional support. A.R.-F. was supported by the FIAS fellowship Program, co-funded by the European Commission, Marie-Skłodowska-Curie Actions - COFUND Program, grant no 945408. P.L.-C. was funded by Ministerio de Universidades of the Spanish Government (Margarita Salas para la formación de jóvenes doctores). The graphical abstract contains an icon that was generated in BioRender. Author contributions Conceptualization, P.L.-C. and A.R.-F.; methodology, P.L.-C. and A.R.-F.; software, P.L.-C. and S.W.; formal analysis, P.L.-C., A.P.-C., and S.W.; investigation, P.L.-C. and A.P.-C.; resources, P.L.-C. and A.P.-C.; data curation, P.L.-C. and S.W.; visualization, P.L.-C.; project administration, P.L.-C. and A.R.-F.; writing – original draft, P.L.-C. and A.R.-F.; writing – review and editing, P.L.-C., A.P.-C., T.F.M., S.W., and A.R.-F.; supervision, A.R.-F.; funding acquisition, A.R.-F. Declaration of interests The authors declare no competing interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERDeposited dataDataThis paperhttps://doi.org/10.17605/OSF.IO/AVY4WSoftware and algorithmsCustom codeThis paperhttps://doi.org/10.17605/OSF.IO/AVY4WMATLAB R2019bhttps://www.mathworks.com/products/matlab.htmlRRID: SCR_001622EEGLAB v14.1.2.https://sccn.ucsd.edu/eeglab/RRID: SCR_007292ERPLAB toolbox v7.0.0.http://erpinfo.org/erplabRRID: SCR_009574Mass Univariate Toolboxhttps://web.archive.org/web/20250214093450/http://openwetware.org/wiki/Mass_Univariate_ERP_ToolboxRRID: SCR_016108G∗Power v3.1.9.7https://www.psychologie.hhu.de/arbeitsgruppen/allgemeine-psychologie-und-arbeitspsychologie/gpowerRRID: SCR_013726R version 4.5.2.https://www.r-project.org/RRID: SCR_001905Python v3.1.3.http://www.python.org/RRID: SCR_008394R package ‘lme4’https://cran.r-project.org/web/packages/lme4/index.htmlRRID: SCR_015654 Experimental model and study participant details Fifty right-handed, young healthy adults participated in the experiment in exchange for monetary compensation. All participants were native Spanish speakers (Spanish monolinguals or Spanish-Catalan bilinguals). Four participants were excluded from the study due to technical problems. In addition, one participant was excluded after EEG data pre-processing, thus leaving a final sample of 45 participants for the final data analyses (28 women; M = 21.4 years; SD = 2.6 years). No participants were excluded based on behavioral performance. As a check against attentional disengagement, we verified that only one participant showed below-chance performance, exclusively in the uninformative condition, a pattern more consistent with condition-specific difficulty than with global disengagement from the task (see Figure S1). Since the effect of participant sex was not central to the study, we did not recruit a balanced sample of males and females (17 males, 28 females) and therefore statistical power was insufficient to reliably assess for sex-related differences. Future research should recruit larger and more balanced samples to address this limitation. The experiment was approved by the Biomedical Research Institute of Bellvitge (IDIBELL) and Hospital Universitari de Bellvitge ethics committee (CEIC, Ref. PR254/15). All participants gave written informed consent before participating in accordance with the principles of the Declaration of Helsinki. Method details Materials and design Sentence triplets We set up a contextual word learning task in which participants had to learn the meaning of a new word from three-sentence contexts (Table 1). Broadly, the first sentence introduced the new word (hereafter referred to as “target item”), the second sentence contributed different amounts of information about its meaning, and the third sentence revealed its superordinate term (see Figure 1). Examples of sentence tripletsExampleaConditionSentence 1Sentence 2Sentence 31InformativeHe likes to walk in the ánaro. (Suele salir mucho a pasear en ánaro).bHe enjoys it when the weather’s hottest. (Disfruta de las horas de más calor.)It’s his favorite season of the year. (Esa es su estación favorita del año.)UninformativeHe likes to walk in the ánaro. (Suele salir mucho a pasear en ánaro).He has always been great with numbers. (Se le dan muy bien los números.)It’s his favorite season of the year. (Esa es su estación favorita del año.)ControlHe likes to walk in the summertime. (Suele salir mucho a pasear en verano).He enjoys it when the weather’s hottest. (Disfruta de las horas de más calor.)It’s his favorite season of the year. (Esa es su estación favorita del año.)2InformativeHis father was very good at parlontas. (Su padre es bueno en las parlontas.)He has always been great with numbers. (Se le dan muy bien los números.)He thought other subjects were very boring. (En las demás asignaturas se aburría mucho.)UninformativeHis father was very good at parlontas. (Su padre es bueno en las parlontas.)He enjoys it when the weather’s hottest. (Disfruta de las horas de más calor.)He thought other subjects were very boring. (En las demás asignaturas se aburría mucho.)ControlHis father was very good at mathematics. (Su padre es bueno en las matemáticas.)He has always been great with numbers. (Se le dan muy bien los números.)He thought other subjects were very boring. (En las demás asignaturas se aburría mucho.)aA given participant would be presented with only one of the sentence triplets in the example.bOriginal Spanish sentences are in brackets. English translations are approximate, as they have been adapted to fulfill the criteria that each sentence must have seven words. Examples of sentence triplets A given participant would be presented with only one of the sentence triplets in the example. Original Spanish sentences are in brackets. English translations are approximate, as they have been adapted to fulfill the criteria that each sentence must have seven words. The first sentence introduced the target item (i.e., Spanish pseudoword) in its final position (“Nunca ha sido alérgico a las nilanas”, in English, “She has never been allergic to nilanas”). The first sentence did not provide enough information to elucidate the word’s meaning, as confirmed by the first norming study (see Norming studies section). However, the combination of the first and the second sentence provided either an uninformative context (Uninfo) (‘She found it in her neighbor’s garage.’) or an informative context about the meaning of the target item (‘She picked one from her own garden.’). The difference in the level of informativeness between the two conditions was corroborated by the second norming study. Importantly, the second sentence was counterbalanced between the two conditions (Info and Uninfo), such that the same sentence set an informative context for some participants, and an uninformative context for others. This was done to control for differences in syntactic structure, as explained in more detail below. Finally, the third sentence revealed the superordinate concept (‘flowers’), which always appeared at the fourth position of the sentence (‘Many of the flowers are newly planted’). The preceding three words (‘Many of the … ’) were always function words (articles or prepositions but never nouns or verbs) that did not provide semantic information about the meaning of the target item or its superordinate category (Table 1). In addition to the learning conditions (Info and Uninfo), we included control contexts (Control condition) in which participants were not required to learn because the target item was a known, real word. Therefore, the remaining context was informationally redundant about the target item’s meaning. In this condition, sentence triplets were the same as those in the Informative and Uninformative conditions except for the target item (of note, participants only encountered each sentence triplet once in the experiment). To create the sentence triplets, we selected 48 superordinate concepts (e.g., flower) and 3 subordinate concepts (e.g., rose, tulip and daisy), thus obtaining a total of 144 subordinate concepts. Then, each subordinate concept was assigned a pseudoword that followed Spanish phonotactic rules. Pseudowords were characterized using several psycholinguistic metrics (for methodological details, see Figures S2, S3, and S4; Table S1). Summed phonotactic probability at both the phoneme-level (pseudowords: M = 0.57, SD = 0.12; real words: M = 0.52, SD = 0.18) and biphone-level (pseudowords: M = 0.05, SD = 0.02; real words: M = 0.05, SD = 0.03) was comparable between pseudowords and real words, indicating that pseudowords followed typical Spanish phonological patterns. Orthographic neighborhood size was lower for pseudowords, as expected (pseudowords: M = 1.53, SD = 2.64; real words: M = 6.82, SD = 8). Moreover, pseudowords had fewer phonological neighbors than real words (pseudowords: M = 2.25, SD = 3.56; real words: M = 8.41, SD = 9.76). All pseudowords differed from their real words by at least three-character operations (orthographic Levenshtein’s distance: M = 6.02, SD = 1.37). Finally, all the pseudowords can be transformed into a real word within three-character operations (orthographic Levenshtein’s distance: M = 1.51, SD = 0.58, range = 1–3). Critically, orthographic and phonological properties of the pseudowords did not vary systematically across conditions or counterbalancing lists. As nouns have gender class in Spanish (masculine or feminine), the pseudoword was gender-matched to the Spanish word corresponding to subordinate concept to avoid syntactic incongruence.53 For instance, if the real Spanish word corresponding to the subordinate concept was “rosa” (feminine), the gender-matched pseudoword “nilana” was used. Furthermore, pseudowords were created with the requirement that they differed in more than one syllable from real Spanish words. The correspondence between pseudowords and subordinate concepts was the same for all participants. Then, we created 3 sentence triplets for each subordinate concept (e.g., ‘rose’), one per experimental condition (Info, Uninfo and Control). As previously explained, a crucial feature of the design of the sentence triplets was that the second sentence was counterbalanced across conditions and across participants, with the goal of controlling for any effects associated with differences in the syntactic structure complexity. To achieve this, every subordinate concept was paired with another subordinate concept from a different superordinate category, for example, “rose” was paired with “motorbike” (i.e., from the superordinate categories “flower” and “vehicle”, respectively). After pairing the terms, we generated their contexts side-by-side such that the same second sentence could be plausibly used in both cases, while at the same time establishing different levels of informativeness. We controlled that the subordinate concepts within the same superordinate category (e.g., rose, tulip and daisy) were paired with subordinate concepts from different superordinate categories (e.g., rose-motorbike [flower-vehicle], tulip-ring [flower-jewellery] and daisy-dolphin [flower-animal]). This was done to avoid potential implicit or explicit associations between superordinate categories during the experiment, for example, between “flower-like” and “vehicle-like” sentence triplets, which could influence learning. To counterbalance the materials, we generated three different lists such that each subordinate concept appeared under a different condition in every list (e.g., the subordinate concept “rose” appeared in the Info condition in list 1, in the Uninfo condition in list 2, and in the Control condition in list 3). Each participant was exposed to one of the lists. The subordinate concepts appeared in the same experimental block in every list, for example, “rose” always appeared in the first experimental block for all participants (albeit in different conditions). Finally, we used natural language processing methods to quantify the syntactic complexity and semantic similarity of sentence triplets and compared them across conditions (for details, see Figures S5 and S6). Two corpus-based measures of syntactic complexity –mean dependency distance and dependency relation frequency– did not differ across conditions, indicating comparable syntactic properties. Regarding semantic similarity, the second sentence was semantically more related to the first and third sentences in the Info and Control conditions than in the Uninfo condition. This is an expected result, considering that Uninfo contexts were created by having a second sentence that was semantically distant to the target word. Norming studies To quantify the level of informativeness of the sentence triplets, we conducted two norming studies. In these studies, we computed the cloze-probability54 of the subordinate concept within the sentence triplet, i.e., the percentage of participants that used the subordinate concept to complete the sentence triplet (for a similar approach, see Mestres-Missé et al.4). In the first norming study, sixteen participants (8 women; mean age = 22.4 ± 3.4 years) read the first sentence of the 144 sentence triplets with the final word missing (e.g., ‘She has never been allergic to the’]) and were asked to “fill the gap with the first word that came to their mind and that fit the sentence”. As intended, when only the first sentence was provided, the cloze-probability of the subordinate concept was very low (mean cloze-probability = 8.85%). In the second norming study, a group of 24 participants (10 women; M = 27.1 years, SD = 2.1 years) read the first and the second sentence of each triplet, with the target item missing (e.g., ‘She has never been allergic to the _. She found it in her neighbor’s garage.’) and to “fill the gap with the first word that came to their mind and that fit the sentence”. To prevent the same participant from reading the same second sentence under different conditions (Info and Uninfo), we divided the sentence triplets in two counterbalanced lists of 144 sentence triplets (48 Info items and 48 Uninfo items), and participants were assigned to one or the other. After collecting the data, we computed two measures for each sentence triplet: 1) the mean cloze-probability of the expected word for the subordinate concept (same-word criteria, e.g., ‘rose’) and 2) the mean cloze-probability of the expected superordinate category (same-category criteria, i.e., any word alluding to a subordinate concept from the superordinate category ‘flower’). Using the same-word criteria, the mean cloze-probability ±SEM was 43.3% ± 2.28% in the Info condition and 2.3% ± 0.5% in the Uninfo condition, t(286) = 17.56, p <.001. Using the same-category criteria, the mean cloze-probability ±SEM was 63.9% ± 2.15% in the Info condition and 10.6% ± 1.4% in the Uninfo condition, t(286) = 20.75, p <. 001. Therefore, as expected, the behavioral tests confirmed that the two conditions (Info and Uninfo) differed in the level of informativeness about the meaning of the target item –operationalized as mean cloze-probability. Also, sentence triplets were more effective in mapping the broader superordinate category than the specific subordinate concept, both in the Info condition (mean cloze-probability same-category = 63.9%; mean cloze-probability same-word = 43.3%), t(286) = 6.59, p <.001) and in the Uninfo condition (mean cloze probability same-category = 10.6%; mean cloze probability same-word = 2.27%), t(286) = 5.402, p <.001. Forced-choice recognition test Finally, to measure the participants’ learning success, they were tested on the meaning of the target items through a forced-choice test after every experimental block (i.e., every 18 triplets; see Procedure section for details). The target item was presented along with four possible responses: the correct meaning (“rosa” [rose]) and three foils (“termómetro” [thermometer], “academia” [academy] and “actriz” [actress]) (chance level = 25%). The four choices (correct meaning and foils) were randomly ordered and remained on screen until the participant responded. All foils were words from a different superordinate category to that of the correct meaning. For example, if the meaning of “nilana” was “rose”, then none of the foils belonged to the superordinate category “flower”. In addition, the foils were assigned such that two of them were gender-matched with the target item (i.e., “academia” and “actriz” are feminine, like “nilana”). This was done to prevent participants from strategically using the gender to guess the correct response. One of the foils was the meaning of another target item that had been presented under the same condition in the previous block. In the current example, the foil “termómetro” [thermometer] was the meaning of the target item “pistano”, which had also been presented under the Informative condition. The other two foils were new words that did not belong to any task-relevant superordinate categories (“academia” [academy] and “actriz” [actress]). A list of 96 new words was elaborated, matched with the full list of 144 Spanish words of the subordinate concepts in terms of mean word frequency per million obtained from EsPal.55 During the experiment, each word appeared twice in the forced-choice test, ensuring an equal exposure to all words. In the case of words of the subordinate concepts, they were presented once as the correct choice and once as an incorrect choice. Procedure Before starting the experiment, participants signed the informed consent and were seated comfortably in front of the computer screen leading to a viewing distance of 70 cm. Then, the EEG cap was placed on the participant’s scalp and each electrode was checked and adjusted until all impedances were below 5 kΩ and the signal was optimal. Before initiating the experiment, participants were given the following instructions: “You will read sentence triplets (that is, groups of three sentences) that will be presented one word at a time. Sometimes, the first sentence will end with an invented word. Your goal is to discover its meaning through the information provided by the second and third sentence that follow immediately after. It is important that you do your best to find out its meaning because you will be asked about it at the end of each block.” They were encouraged to minimize movements and to refrain from blinking until the blinking signal appeared at the end of each trial. Participants read 144 sentence triplets presented in 8 blocks of 18 sentence triplets (Figure 1). Every block contained 6 sentence triplets of each experimental condition (6 Uninfo, 6 Info and 6 Control), which were presented in randomized order. Sentence triplets were presented as follows (Figure 1B). First, a fixation cross appeared at the center of the screen for 1000 ms. Then, the first sentence was presented one word at a time (white background, black monospaced font; font: Courier New, 12 pts font size). Each word remained on screen for 200 ms, followed by a 335 ms inter-word interval. The last word of the first sentence (i.e., target item) was presented for 700 ms, followed by a 335 ms blank screen. After the blank screen, the blinking signal (a small circle) appeared at the center of the screen, indicating participants that they could blink. The blinking signal lasted 2500 ms. Then, after a 1035 ms blank screen, the second and third sentences were presented one word at a time (200 ms + 335 blank screen ms for all words). The second and third sentences were separated by a 735 ms blank screen. At the end of the third sentence, the blinking signal appeared again for 2500 ms before the next sentence triplet began. After completing a block (i.e., 18 sentence triplets), participants performed the forced-choice test about the meaning of the new words in that block (Figure 1B). Thus, they were tested on a total of 12 new words per block (6 Info and 6 Uninfo), presented one at a time and in random order. The new word was displayed at the top of the screen, and the four choices were shown below in different lines (with interspace) and numbered (1–4). This information was displayed until participants made their choice by pressing a key on the keyboard (1, 2, 3 or 4). Right after that, the chosen word changed its color to provide feedback on whether their response was correct (green) or incorrect (red). Thus, participants were not given feedback about the correct meaning of the word, but only about the correctness of their response. This was done to encourage participants to rely on their knowledge and inferences during the test, instead of simply guessing and waiting for the feedback. EEG data acquisition and pre-processing Electroencephalographic data (sampling rate = 250 Hz; on-line bandpass filter = .015–250 Hz) were recorded from 32 tin scalp electrodes at standard 10/20 system positions (electrode positions: FP1/2, Fz, F3/4, F7/8, FC1/2, FC5/6, Cz, C3/4, CP1/2, CP5/6, Pz, P3/4, T3/4, T5/6, PO1/2, O1/2, left and right mastoids). Vertical and horizontal electro-oculograms were recorded and used for artifact rejection purposes. All electrode impedances were kept below 5 kΩ during the session. EEG data were pre-processed with the ERPLAB toolbox v7.0.056 integrated in EEGLAB v14.1.257 and running under MATLAB R2019b. Customized MATLAB scripts were used for automatizing the pre-processing steps and for extracting the structured data values for statistical analyses in R. The continuous data were re-referenced offline to the mean activity of the left and right mastoid electrodes. A notch filter at 50 Hz was applied to reduce electrical line noise. Then, the continuous data recordings were segmented into epochs time-locked to the following events: 1) target item (from −200 to 735 ms, time-locked to the onset of the last word of the first sentence), and 2) second and third sentences together (from −200 to 8225 ms, time-locked to the onset of the first word of the second sentence). Artifact detection and rejection were applied on the segmented data. For the target item interval, the epochs were baseline-corrected to the 100 ms pre-stimulus interval and then screened for artifacts using a simple voltage threshold, whereby epochs with voltage amplitudes ±100 μV at the ocular channel (vertical electro-oculogram) or ±200 μV at any other channel were rejected. Applying this criterion, an average of 95% trials per condition remained available for analysis (Info = 45.6 trials; Uninfo = 45.3 trials; Control = 45.7 trials). For the second and third sentence interval, the concatenated epochs were subjected to independent component analysis (ICA) using the “runica” algorithm implemented in EEGLAB and excluding the ocular channels for the decomposition. Then, the resulting components were visually inspected and those corresponding to artifacts were rejected. The number of rejected components per participant spanned from 1 to 4 components (mean = 1.8 components, SD = 1). Noisy or unstable electrodes were interpolated on fourteen participants using spherical interpolation as implemented in ERPLAB. A maximum of 2 electrodes were interpolated per participant. After artefact-correction, epochs were baseline-corrected to the 100 ms pre-sentence onset interval and were rejected if they contained voltage amplitudes ±100 μV at the ocular channel (vertical electro-oculogram) or ±200 μV at any other channel. An average of 93% trials per condition were available for analysis (Info = 44.7 trials; Uninfo = 44.2 trials; Control = 44.4 trials). Participants with less than 30% available trials for any epoch of interest were excluded from further analyses, resulting in 1 excluded participant. Finally, the epochs were averaged for each participant and condition, and the resulting averaged ERP waveforms were visually inspected to detect any remaining participant with excessively noisy data (i.e., large drifts, muscular activity, or unstable channels). No participant was excluded in this step. Quantification and statistical analysis Event-related potential analyses For each condition, ultraslow activity was computed in the interval from the onset of the second sentence to the end of the sentence triplet (i.e., 8225 ms interval), with a 100 ms pre-stimulus interval (León-Cabrera et al., 2019). Importantly, to isolate ultraslow modulations from concurrent word-elicited responses at higher-frequencies, we applied a low pass filter at 0.7 Hz (Butterworth filter with a roll-off of 12 dB/oct) which has proven effective in capturing ultraslow modulations during sentence processing.35,39 The same filter was used for the figures presented. To evaluate differences in the amplitude of ultraslow modulations between conditions, we applied cluster-based permutation tests58 as implemented in the Mass Univariate Toolbox59 running under EEGLAB. In brief, these tests identify adjacent time points and/or electrodes with similar differences between conditions while successfully controlling the family-wise error rate (FWER). Specifically, a t-value is obtained for every time point and electrode for each contrast (i.e., Info minus Control; Uninfo minus Control, and Info minus Uninfo) through paired samples t-tests. Clusters were defined as nearby data points with t-values above a threshold of 2.01 (alpha level of .05 and with 44 degrees of freedom, for two-tailed testing), and a cluster-level statistic was obtained by summing up the t-values within every cluster and retaining the value of the maximum cluster-level statistic (observed cluster statistic). To determine if there were significant differences between the two conditions, the prior steps were repeated on random partitions of the data (10,000 permutations). Lastly, the distribution of the cluster-level statistic of the random partitions was used to derive a p-value, defined as the proportion of random partitions that yielded a cluster-level statistic larger than the observed cluster statistic (alpha level of .05). To assess N400 effects at the superordinate word (i.e., fourth word of the third sentence), we ran cluster-based permutation tests between every condition: Uninfo minus Info (i.e., informativeness effect), Uninfo minus Control, and Info minus Control (alpha level of .05 and with 44 degrees of freedom, for two-tailed testing; 5,000 permutations). In this case, the analysis was constrained in the temporal domain by averaging mean amplitudes in the 300 to 500 ms post-word onset interval (100 ms pre-stimulus baseline) where N400 effects were expected. As a control analysis, we computed N400 effects using the same parameters at the target item (i.e., last word of the first sentence) to ensure that new words were processed equally in both learning conditions (Info vs. Uninfo) and that, at the same time, these elicited distinct N400 responses from real words. The data were low pass filtered at 30 Hz for the analyses and at 12 Hz for the figures presented (Butterworth filter with 12 dB/oct roll-off). Relation between learning performance and electrophysiological components To assess the relation between electrophysiological and behavioral measures, we ran linear mixed-effects (LME) models (using the ‘lme4’ package60) and correlations between electrophysiological components (ultraslow modulations and N400 effects) and learning performance, quantified as described next. Regarding the correlational analyses, a sensitivity analysis using G∗Power61 to determine the minimum effect size detectable with 80% power given our sample size (N = 45) and alpha level (α = .05, two-tailed), revealed that our study was adequately powered to detect correlations of r ≥ .40. In addition, to assess the reliability of the correlation estimates, we computed 95% percentile bootstrap confidence intervals (5000 iterations) following,62 as implemented in the boot package in R.63 All correlations were were corrected for multiple comparisons with the Holm-Bonferroni correction (as implemented in the ‘rcorr.adjust’ function). Corrected p-values are reported as pHB. Learning performance measures Learning performance was quantified as the mean percentage of correct responses in the forced-choice test responses for every participant with Rstudio (version 1.4.1106). We distinguished between overall learning performance (averaged correct responses across learning conditions) and condition-specific learning performance (average of correct responses in the Info or Uninfo condition, individually). Electrophysiological measures All ultraslow modulation measures were computed at a centro-parietal electrode cluster (including Pz, P3, P4, CP1, CP2 Cz, C3 and C4) based on the spatial distribution of the effects observed at the sentence-level, and within the second sentence (from the onset of the sentence until its penultimate word). We did not include the interval after the last word of the sentence to avoid including overlapping sentence-final effects. N400 mean amplitudes were computed in the 300–500 ms interval post-word onset in the same centro-parietal electrode cluster (Pz, P3, P4, CP1, CP2, Cz, C3 and C4) based on the expected distribution of the component.6 Correlations between condition-specific learning performance and ultraslow modulations We captured the swiftness of the build-up of the ultraslow modulations through the computation of their slope within the second sentence. Specifically, the linear slope of the ultraslow modulation was computed by fitting a linear model,64 using the “lm” function in Rstudio (version 1.4.1106) for every participant and learning condition (Info and Uninfo) with mean voltage amplitude in μV as the outcome variable and time in milliseconds as the predictor variable (ranging from 0 to 3210 ms time-locked to the onset of the second sentence). The mean voltage amplitudes were computed at a centro-parietal electrode cluster (including Pz, P3, P4, CP1, CP2, Cz, C3 and C4) based on the spatial distribution of the effects observed at the sentence-level. We obtained two slope measures per participant (one per learning condition). We performed Pearson correlations (with an alpha level of .05) to assess whether the slope of the ultraslow modulation correlated with learning performance in each condition (Info and Uninfo condition separately). Linear mixed-effects (LME): Relation between overall learning performance and ultraslow modulations To complement the correlational analyses, we ran a model to further examine the trajectory of the ultraslow modulation on a word-by-word basis. Instead of reducing the ERP signal to a pre-computed slope, this model included word-level time intervals as a continuous predictor, capturing the unfolding signal at a finer temporal resolution. Furthermore, this analysis used overall learning performance (i.e., average of correct responses across learning conditions), thereby allowing us to test whether the relation generalized to general learning ability in the task. We fitted an LME on the mean amplitude averaged over the centroparietal cluster across the six word-level time intervals (535 ms duration each). A linear time predictor was derived from an orthogonal polynomial contrast over word onsets. The model included fixed effects of Time, Overall Learning (z-scored across participants), and Condition (Info and Uninfo; sum-to-zero coded), along with all interactions. A maximal model with full random-effects structure was constructed (Bar et al., 2013), which included by-participant random intercepts and random slopes for all within-subject fixed factors. We used the Satterwhaite method estimate p-values.65 Follow-up tests on significant interactions were conducted with the ‘emmeans’ function with Holm-Bonferroni correction for multiple comparisons. Correlations between condition-specific learning performance and N400 effects N400 mean amplitudes were computed in the 300–500 ms interval post-word onset at a centro-parietal electrode cluster (Pz, P3, P4, CP1, CP2, Pz, C3 and C4). For every participant, mean amplitude differences were correlated with learning performance (i.e., percentage of correct responses) in each learning condition (Info and Uninfo condition). We computed Pearson correlations between learning performance and the following N400 effects at the superordinate word: Uninfo minus Info, Control minus Info, and Control minus Uninfo. Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.isci.2026.117159. Contributor Information Patricia León-Cabrera, Email: pleoncabrera@ub.edu. Antoni Rodríguez-Fornells, Email: antoni.rodriguez@icrea.cat. Supplemental information Document S1. Figures S1–S7 and Table S1 References Associated Data Supplementary Materials Document S1. Figures S1–S7 and Table S1 Data Availability Statement •The anonymized EEG data have been deposited at Open Science Framework (OSF) and are publicly available at https://doi.org/10.17605/OSF.IO/AVY4W as of the date of publication. DOIs are also listed in the key resources table. The anonymized EEG data have been deposited at Open Science Framework (OSF) and are publicly available at https://doi.org/10.17605/OSF.IO/AVY4W as of the date of publication. DOIs are also listed in the key resources table. •All original code used for EEG data preprocessing and analysis have been deposited at OSF and is publicly available at https://doi.org/10.17605/OSF.IO/AVY4W as of the date of publication. DOIs are also listed in the key resources table. All original code used for EEG data preprocessing and analysis have been deposited at OSF and is publicly available at https://doi.org/10.17605/OSF.IO/AVY4W as of the date of publication. DOIs are also listed in the key resources table. •Any additional information required to reanalyze the data in this article is available from the lead contact upon request. Any additional information required to reanalyze the data in this article is available from the lead contact upon request.