Dynamics of the attentional blink in preverbal infants

Significance

The global workspace (GWS) enables the maintenance of stable, explicit representations that can be flexibly accessed by different cognitive systems. While infants’ sensory abilities are well established, it remains unclear when such higher-level integrative processing emerges. Using trial-level behavioral and electroencephalographical (EEG) markers, we show that several core properties of the GWS—serial access, attentional competition, and dynamic prioritization—are already functional by 4 mo of age. Although access is slower than in adults, infants can selectively allocate attention both in space and time, retain task-relevant information, and exhibit bottlenecks indicative of central access limitations. These findings suggest that the architecture supporting conscious-level processing emerges earlier than previously thought, bridging perception with nascent cognitive control.

Abstract

In the first months postterm, human infants are traditionally viewed as passive to their environment, unable to focus or sustain attention on specific objects. Here, we asked whether 4-mo-old infants could engage their attention strongly enough in a visual task to block the perception of a subsequent interesting stimulus—a phenomenon known as attentional blink, which in adults reflects a serial processing bottleneck for accessing a central workspace. We presented three successive visual events: a central teddy bear (T1) followed by a lateralized face and scrambled face (T2) at varying stimulus onset asynchrony (SOAs: 400, 800, and 1,200 ms), then the same face reappeared after 1,600 ms (T3). We monitored saccades toward the face, pupil size, and electroencephalic (EEG) responses to the flickering background. Our behavioral and EEG results showed that infants missed the lateralized face at the shortest SOA (400 ms). At 800 ms, detection occurred only when the face appeared in the left hemifield, while detection in the right hemifield required 1,200 ms, suggesting a hemispheric asymmetry in face processing. Furthermore, at SOAs where T2 should be visible, a trial-by-trial metric based on event-related variability revealed that the depth of attentional engagement to T1 predicted access to T2, despite identical visual input. These findings support the presence of a global workspace in early infancy, though with slower dynamics.

For centuries, infants’ mental lives were thought to be empty or confused, dominated by fleeting impressions without the capacity to direct attention toward specific objects or maintain information over time. In adults, such sustained and selective processing is supported by a global workspace (GWS), a central system that enables information to be retained, integrated across modalities, and selectively amplified depending on task goals (1, 2). These functions rely on reverberant activity between the prefrontal and posterior parietal cortices—regions known to undergo prolonged postnatal maturation. For this reason, the early presence of a functioning GWS in infancy was long considered unlikely.

Yet, recent brain imaging studies challenge this view, revealing structural and functional continuity between infant and adult brains. Long-range anteroposterior functional connectivity—supporting the frontoparietal, dorsal attention, and default mode networks—is already observed from the neonatal period (3), and sensory responses in infants propagate through the entire bottom–up hierarchy, reaching high-level integrative regions already before term age, much earlier than previously thought (4–6).

However, more direct evidence is needed to determine whether infants truly benefit from a GWS. One of its defining characteristics is its serial access, which accounts for the fact that processing a first event can momentarily block awareness of a second occurring in close succession—a phenomenon called attentional blink (7). This processing bottleneck stands in contrast to the massively parallel processing at lower-level processing stages and is thought to mark access to a central, capacity-limited processing stage (8).

Two recent studies have tested for such a bottleneck in infancy using rapid serial visual presentation. In one study, 8-mo-old infants saw two faces presented 200 or 800 ms apart within a 10 Hz stream of images (1 image/100 ms) and familiarity was later tested by pairing each face with a novel one (9). A novelty effect was observed for both faces at long SOA (800 ms), but not for the second face at 200 ms- SOA, suggesting a difficulty in encoding the second face, possibly related to an attentional blink or to low-level perceptual masking induced by the rapid presentation rate. Hochman and Kouider (10) used a different design, presenting scrambled faces in three parallel visual streams, at 200 or 300 ms per image. Occasionally, a face appeared centrally, followed—after a variable SOA—by a lateralized face. Tracking saccades to the lateralized face, they described an attentional blink lasting up to 1,200 ms at 5 mo, decreasing with age to ~200 ms by 3 y, close to adult values.

While both studies suggest a temporal processing bottleneck, face detection may rely on automatic feed-forward mechanisms, especially given the strong attentional capture by faces (11) and the risk of backward masking in rapid presentations near infants’ visibility thresholds (12). These factors complicate the interpretation in terms of access to a GWS.

Here, we reexamined this question using EEG with the aim of disentangling early sensory masking effects from a potential higher-level bottleneck caused by serial access to the GWS in younger, 4-mo-old infants (n = 19). We designed a task involving three successive, naturally engaging visual events (Fig. 1). Each trial began with a novel central image of a “teddy bear” (T1) to capture attention. After a variable SOA, a face (T2) appeared in one hemifield, paired with a scrambled face in the opposite hemifield. Finally, the same face (T3) reappeared at the same location, encouraging infants to anticipate upcoming lateralized faces. Trials without T2 were included to compare EEG and saccadic behavior in the presence vs. absence of T2. We hypothesized that infants would exhibit an attentional blink —i.e. a failure to perceive T2— at short, but not long, SOAs.

Fig. 1.

Trial structure. A colored teddy-bear was displayed at the screen center (T1), followed by a face and a scrambled face at three different SOAs (T2). No-T2 trials were also presented. Finally, a face (T3) appeared on the same side than the T2 face 300 ms after the offset of the teddy bear (fixed T2-T3 SOA = 1600 ms). Salt-and-pepper noise patterns were flickering at 10 Hz on each side of the screen throughout the session.

To assess T2 perception, we contrasted trials in which T2 was presented vs. omitted across SOAs and combined behavioral and electrophysiological measures to capture both overt and covert orienting responses. Overt orientation to T2 was assessed with eye-tracking using saccades as well as gaze-orientation accuracy toward the face vs. the scrambled face to distinguish true face recognition from automatic orienting (11). Covert orientation was evaluated through pupil size changes (13) reflecting attentional engagement, and through suppression of steady-state visual evoked potentials (SSVEP), which indicates focused attention on the lateralized image (14, 15) that masks the flickering background (16). Finally, we analyzed how T2 perception might modulate the anticipatory signal (analogous to the adult Contingent Negative Variation or CNV) associated with the expectation of a lateralized face after T1 (17, 18).

Moreover, in addition to comparing T2 perception across SOAs, we analyzed how attentional engagement with T1 influenced T2 perception at the same long SOA. This within-SOA comparison is critical to confirm that missing T2 reflects an attentional blink rather than low-level masking. Since infants cannot provide explicit reports, we applied a framework that treats trial-by-trial neural fluctuations as meaningful indicators of internal processing (19). According to this approach, each stimulus perturbs ongoing neural activity, nudging it toward a characteristic attractor pattern. The speed and trajectory of this convergence reflect the efficiency of stimulus processing—a dynamic we refer to as “fly-by.” Event-related potential (ERP) components provide reliable proxies for these attractor states, as they capture the most consistent spatial patterns of neural activation within specific time windows (19). We focused on the P400 component, a known marker of visual processing in infants, and extracted, for each trial, the latency at which neural activity after T1 came closest to the P400 topography. These fly-by latencies index how quickly and efficiently the infant’s brain responds to T1. Additionally, we examined variability quenching, the reduction of trial-to-trial neural variability that typically accompanies attentional engagement (19, 20). These two measures allowed us to assess how fluctuations in attention to T1 modulated the likelihood of perceiving T2 at a constant, a-priori visible SOA, offering a sensitive window into the dynamics of attentional selection in the infant brain.

Results

Impact of T1-T2 SOA on T2 Perception.

Ocular saccades toward the lateralized face.

Despite competition between T2 images and the central T1 image, a few saccades toward the lateralized images (<30% of trials) were present. We observed a main T2 effect on the number of saccades (binomial LME, 95% CI: [0.42, 1.11]; P < 0.001) but no significant interaction T2 X SOA (SI Appendix, Fig. S3C). No preferred orientation for the face vs. the scrambled face was observed, except a marginal preference for faces at 1200 ms SOA [400 ms: t(18) <1; 800 ms: t(18) <1; 1200 ms: t(18) = 1.7, P = 0.05, one-tailed; Fig. 2B]. This reveals a part of automatic gaze orientation when a faint change occurs in the peripheral visual field, with more voluntary control toward faces emerging at the longer 1200 ms SOA.

Fig. 2.

T2 processing across SOA and T2 conditions. (A) Pupil size response; (B) Saccade performance (orientation to the face); (C) SSVEP. Significant effects are present only for the longest SOA. *P < 0.05.

During the trials where the gaze remained centered, we measured the dilation of the pupil. Across all SOAs, pupil dilation was greater in T2 trials than no-T2 trials between 1000 to 1600 ms after T2 onset (cluster-based permutation, Cohen’s d = 0.64, P = 0.02). Further analyses of this time-window showed a significant T2 X SOA interaction (LME, 95% C.I. [0.01, 0.10], P = 0.01), with larger pupil dilation for T2 vs. no-T2 trials only in the 1200 ms-SOA condition (400 ms: t(18) <1; 800 ms: t(18) = 1.7, P = 0.11; 1200 ms: t(18) = 2.6, P = 0.02; Fig. 2A).

SSVEPs when gaze remained centered.

When merging all SOA conditions, a significant decrease in SSVEP amplitude was observed between 300 to 800 ms after T2 onset in T2 trials compared to no-T2 trials (cluster-based permutation analyses, Cohen’s d = −0.7, P = 0.02). During this interval, a significant T2 × SOA interaction was found (LME, 95% CI: [−0.47, −0.04], P = 0.02), with a significant decrease only in the 1,200 ms SOA condition (400 ms: t(18) = −1.1, P = 0.28; 800 ms: t(18) < 1; 1200 ms: t(18) = −3, P = 0.008; Fig. 2C).

Temporal expectation dipole (TED) from T1 to T3.

A slow neural response resembling the adult CNV has been reported in infants anticipating an upcoming stimulus (16–18). We recorded a sustained dipolar activity from T1 to T3 with a topography consistent with these reports (Materials & Method, SI Appendix, Fig. S3). No effect due to T2 was observed at any SOA.

Hemifield differences.

Given previous findings of infants’ difficulty perceiving faces in the right hemifield (21), we examined whether T2 perception was affected by the hemifield of face presentation. There was a main T2 effect on pupil dilation, SSVEPs modulation, and face orientation performance when faces appeared in the left hemifield, thus processed by the right hemisphere (SI Appendix, Fig. S2 and Table S1). Due to the limited number of trials per face side per SOA, we combined the three measures (pupil dilation, SSVEP amplitude drop, and saccade performance) for T2 vs. no-T2 trials into a composite index after z-scoring each measure. This index was significantly larger for faces in the left hemifield compared to the right only at the 800 ms SOA (SI Appendix, Fig. S2D; Left > Right: SOA 400: t(14) < 1; SOA 800: t(18) = 3.2, Pcorr = 0.015; SOA 1200: t(17) < 1, Bonferroni-corrected). This result confirms the right hemisphere (left hemifield) advantage for face processing in infants (21), similar to adults, and suggests that at 800 ms SOA, perceptual strength was higher for left hemifield faces, facilitating entry into the GWS.

To summarize, our results on correct gaze orientation to the face, pupil dilation, and SSVEPs are coherent, with all three measures converging to suggest that infants detected T2 at a long T1-T2 SOA (1200 ms) but not at a short SOA (400 ms). At an intermediate SOA (800 ms), infants noticed faces only when they appeared in the left hemifield, i.e., processed by the right hemisphere.

Responses to T3.

Ocular saccades toward T3 face.

Infants oriented to a side in 88% (±14%) of trials but frequently looked to the empty hemifield (34% ± 10%). No T2 effect was observed across SOAs (SI Appendix, Fig. S1B). The analysis of saccadic sequences across T2 and T3 revealed more repeated than crossed saccades (t(18) = 2.15, P = 0.045). The small number of trials without orientation precluded meaningful analysis of pupil dilation and SSVEPs.

ERP to the lateralized T3 face.

The N290, but not the P100, was modulated by T2 (P100: all F < 1; N290, T2: F(1,18) = 6.9, P = 0.017; SOA and T2 x SOA: both F < 1). Post hoc analyses showed that the N290 was almost absent when T2 was present at 1200 ms SOA (SI Appendix, Fig. S1A; T2 < No-T2; SOA400, t(18) = 1.4, Pcorr = 0.53; SOA800, t(18) < 1; SOA1200, t(18) = 2.8, Pcorr = 0.036, Bonferroni-corrected).

To rule out power-related effects, we regrouped SOA conditions taking into account the 800-ms hemispheric asymmetry. Specifically, trials with a 1200 ms SOA and with an 800 ms SOA and left-hemifield faces were assigned to the high-T2-perception category, whereas 400 ms SOA trials and 800 ms SOA trials with right-hemifield faces were assigned to the low-T2-perception category. The results remained consistent with our previous findings (SI Appendix, Figs. S4 and S5).

In summary, the effect of T2 on T3 was limited, consisting of a reduced N290 when T2 was previously seen and a bias to repeat orientation to the same hemifield as the previous saccade.

Fly-Bys Allow Us to Split Trials Based on Visual Engagement with T1.

To differentiate genuine attentional blink from low-level perceptual limitations, we focused on High-T2 trials in which T2 was a priori visible and quantified T1 processing depth via fly-by latency and variability quenching (VQ) measures. Using the robust P400 following T1, we derived individual P400 templates, extracted each trial’s fly-by latency (i.e., the time of closest approach to the template), and split trials into Early and Late P400 sets based on each infant’s median latency.

The Late P400 set showed a delayed P400 relative to the Early set (Fig. 3A), challenging the notion of a broad, sustained plateau and instead suggesting that the P400 reflects a superposition of temporally variable responses, indexing distinct processing trajectories. Supporting this, a significant posterior negativity—consistent with the N290, typically linked to face perception and attentional engagement in infants (22, 23) —preceded the Late P400 (cluster-based permutation test, [290 to 410 ms], ERP < 0, Cohen’s d = –1.5, P = 0.01). VQ analyses further confirmed distinct temporal dynamics: Trial-to-trial variability decreased earlier in the Early P400 set ([390 to 500 ms], Cohen’s d = –0.9, P = 0.04) than in the Late set ([490 to 620 ms], Cohen’s d = –0.8, P = 0.02), with an additional early quenching in the Late set aligned with the N290 ([310 to 420 ms], Cohen’s d = –1.0, P = 0.02) (Fig. 3B).

Fig. 3.

Attentional engagement to T1 in High-T2 trials, revealed by P400 Fly-bys. (A) ERP at the posterior electrode cluster (see scalp display) for early and late P400 sets of trials; A clear N290 is visible in late P400 trials (see topographic Inset); (B) Variability quenching events in the two sets (*P < 0.05); (C) Predicted attentional modulation patterns across trial events for each set; (D) pupil size at T1 onset (*P < 0.05); (E) TED amplitude in no-T2 trials in both sets.

Further differences between trial sets were observed in other measures. In no-T2 trials, TED tended to be larger in the Early P400 set ([1500 to 2100 ms], Cohen’s d = –0.96, P = 0.08), suggesting a stronger expectation of the upcoming T3 image and, by inference, earlier disengagement from T1 (Fig. 3E). Finally, pupil size was significantly larger in the Late P400 set at T1 onset (F(1,18) = 4.8, P = 0.042), followed by a stronger constriction (Late P400 < Early P400, t(18) = −2.33, P = 0.031), indicating distinct attentional states already present at trial onset, likely influencing the depth of T1 processing (Figs. 3D and 4C).

Fig. 4.

T2 processing (T2 and no-T2 trials) in Early P400 (Top row) and Late P400 (Bottom row) trial sets. (A) TED amplitude; (B) SSVEP modulations; (C) Pupil size response; (D) Orientation to the face (saccade accuracy); (E) Composite T2 perception index combining all measures. (*P < 0.05, **P < 0.01).

The Processing of T1 Influences T2 Perception.

We then examined how visual engagement with T1 affected the behavioral and electrophysiological responses tested in the first section. We predicted that when attention was strongly captured by T1, as indicated by the N290–P400 complex observed in Late P400 set, infants would miss T2, signaling an attentional bottleneck similar to that observed at short T1-T2 SOAs (Fig. 3C).

Early P400 set of trials (shallow T1 processing).

Infants made more saccades when T2 was present (t(18) = 3.9 P = 0.001) and oriented more often toward the face than the scrambled face (Face: 66%, t(14) = 2.2, P = 0.02, one-tailed) indicating selective orienting toward the face over the scrambled stimulus. In trials without saccades, SSVEPs at 10 Hz significantly decreased (T2 < No-T2, t(18) = −3, P = 0.008) indicating covert orientation to T2. TED was reduced after T2 onset, indicating interference with T2 processing (t(18) = 2.9 in [1400, 2000 ms], P < 0.01). However, no significant effect was observed on pupil dilation (T2 > No-T2: Early in [1000, 1400 ms], t(18)<1; Late in [1800, 2400 ms], t(18) = 1.3, P = 0.2).

Late P400 set of trials (deep T1 processing).

None of the previous indicators reached significance. Infants did not make more saccades when T2 was present (T2 > No-T2, t(18) = 1.2, P = 0.12, one-tailed), though there was a marginal tendency to orient toward faces over scrambled faces (61%, t(18) = 1.6, P = 0.06, one-tailed). The T2-related SSVEP decrease ([300 to 800 ms], t(18) = –1.0, P = 0.33) and TED amplitude modulation ([1400 to 2000 ms], t(18) < 1) were also nonsignificant. However, pupil responses revealed both an early and late T2 effect ([1000 to 1400 ms], t(18) = 4.0, P < 0.001; [1800 to 2400 ms], t(18) = 3.4, P = 0.003). However, the late dilation effect was no longer significant after controlling for the early component (t(18) < 1), suggesting a common underlying cause.

Although individual measures pointed toward better T2 perception in Early P400 trials, no significant interaction between dataset and T2 (present vs. absent) was found—likely due to limited trial numbers. To increase sensitivity, we computed a T2 perception Index (T2_I; Fig. 4E), averaging T2 and no-T2 differences across all z-scored measures: saccade rate, saccade performance, pupil dilation, SSVEPs, and TED amplitude (see SOM and SI Appendix, Fig. S4D). These measures reflect complementary, converging indicators of perceptual processing, making their aggregation both principled and interpretable. The resulting T2_I was significantly higher in Early than in Late P400 trials (t(18) = 2.1, P = 0.049), suggesting enhanced T2 perception when attention was only superficially engaged with T1.

Discussion

An Attentional Bottleneck in the Infant Brain.

Our results provide converging behavioral and neural evidence for the presence of an attentional processing bottleneck in 4-mo-old infants replicating and extending prior behavioral finding at later ages (9, 10): When two stimuli (T1 and T2) were presented in rapid succession (400 ms SOA), infants showed no behavioral or neural signatures of T2 detection—no pupil dilation, no increase in face-directed saccades, and no modulation of SSVEP responses. In contrast, at longer SOA (1200 ms), clear signs of T2 processing were present, including increased pupil size, reduced SSVEP power, and more saccades toward faces. Moreover, to ensure this effect was not due to low-level visual limitations or masking, we conducted a within-SOA analysis focusing on High-T2 trials, where T2 was a-priori visible. We leveraged on the temporal variability in P400 responses to classify trials depending on attentional engagement with T1 and showed that failure to process T2 resulted from deeper engagement with the central T1, not from perceptual difficulties. As in adults (8), what matters is not merely the physical delay between stimuli, but the subjective availability of central processing resources—specifically, whether the global workspace has been cleared. These findings reveal a serial processing bottleneck that contrasts sharply with the parallel operations observed at the perceptual level (24, 25).

Our paradigm relied on the competition between a central and a lateralized stimulus—an effective but not unproblematic design. Although peripheral and low-level visual brain regions are well developed by the end of the first trimester (21), and T2 duration used here exceeded the visibility threshold for lateralized faces at this age (12), detecting a lateralized, isoluminant face on a flickering background remains inherently challenging—especially when it competes with a centrally located, highly salient image. Moreover, in adults, reduced N170 amplitudes for lateralized faces presented alongside central faces have been attributed to local competition in visual cortex representations, rather than to failures in GWS access (26). To disentangle these possibilities in infants, it was essential to compare responses under identical visual conditions but with differing levels of attentional engagement to T1. The fly-by methodology enabled this contrast, revealing that impaired T2 perception did indeed stem from a central bottleneck rather than from low-level visual limitations.

Interestingly when we considered both trials with deep (late P400 dataset) and shallow (early P400 dataset) attention to T1, infants showed significant differences in pupil dilation before T1 onset, indicating that they were already in a different attentional state (27). Thus, both the visual features of the stimuli and the infant’s ongoing attentional state played a role in determining their engagement toward the teddy bear and the course of the trial over several seconds until T3 was presented.

A Functional but Immature Global Workspace.

The prolonged interference period (≧800 ms) aligns with results in older infants (10), and with the generally slow orienting responses observed in awake infants to sudden changes (28) or salient stimuli (16, 29, 30). Although this extended interference might be amplified by the challenging visual competition in our paradigm, it likely reflects a more fundamental developmental constraint: slow access to explicit representations (29, 31).

In adults, the GWS is explained by activity in recurrent fronto-parietal networks. While the basic architecture of these long-range networks is already detectable in full-term neonates (3) and infants exhibited frontal activation during spatial attention tasks (6), the frontal and parietal cortices are among the least mature brain regions at this age. Additionally, the slow migration of inhibitory interneurons in prefrontal and cingulate areas (32) may further disrupt GABAergic balance between frontal and parietal regions, an imbalance linked to individual differences in attentional blink magnitude in adults (33). These cytoarchitectonic features likely contribute to the longer-lasting bottleneck observed in infants (10).

Responses to T3 further underscored the immaturity of attentional control in infants. Despite high overall orientation rates, infants frequently saccaded to the incorrect hemifield, repeating a prior orientation made at T2. This pattern may partly reflect perseverative behavior linked to immature frontal inhibitory control, as observed in other domains such as reaching (34). Another contributing factor might be infants’ difficulty integrating face identity (a ventral stream process) with spatial location (computed in the dorsal stream), a known challenge at this age (35). Notably, only trials in which T2 was perceptible (High-T2 trials) modulated the N290 response to T3, despite identical T2–T3 intervals across conditions. The infant brain thus responded to face repetition even when orientation failed, suggesting a possible competition between identity and spatial information within the global workspace, with priority given to identity at the expense of spatial updating.

Taken together, our results provide compelling evidence that core computational principles of the global workspace—serial access, competition for entry and dynamic prioritization, are already operating in the infant brain. Four-month-old infants can maintain task-relevant information over extended temporal windows and flexibly allocate attentional resources. This early-emerging architecture mirrors the adult GWS (1), which enables the active maintenance and reactivation of mental representations as long as they remain relevant (1).

GWS and Temporal Expectation.

This interpretation aligns with prior evidence that infants can form sequential predictions based on structured input. For instance, 5-mo-olds trained to associate specific repetition patterns in speech with particular images exhibit a cascade of neural responses over several seconds including an expectation signal following the word, priming of the visual cortex for the predicted image, and a surprise response when the image was incongruent (16). Similarly, Mento et al. (17) reported a slow, sustained anticipatory signal in infants—akin to the CNV in adults—when awaiting a predictable visual event. In our study, a comparable fronto-occipital dipole gradually built up and resolved at T3 onset, suggesting that infants had learned the temporal structure of the task and were anticipating a lateralized face following the teddy bear (SI Appendix, Fig. S3 A and B). This temporal expectation was particularly visible in no-T2 trials in Fig. 3, where TED amplitude was larger when infants disengaged from T1 earlier—consistent with flexible, goal-directed attention.

However, TED amplitude dropped significantly in early P400 trials where infants attended to T2, compared to no-T2 trials (Fig. 4A). In adults, similar CNV reductions occur in dual-task situations, when distractors disrupt a primary temporal expectation task (36, 37). This suggests that the attentional demands of processing T2 interfered with the temporal prediction of T3. One possibility is that T3 anticipation requires sustained attentional resources, which are diverted when T2 enters the global workspace. Alternatively, detecting T2 may fulfill a general expectation for a lateralized face, prematurely terminating anticipatory buildup. Both interpretations converge on a key point: maintaining temporal predictions requires central attentional resources and is vulnerable to interference from salient events.

Spatial Attention and Face Processing.

Our task encouraged infants to monitor the periphery in anticipation of faces, fostering a broad attentional focus that could dynamically narrow or expand depending on internal expectations or stimulus salience (38). The spatial spread of attention can be tracked through pupil size, which constricts with focused attention and dilates as attention broadens (39), as well as through SSVEP modulation in response to the flickering background (15). These two markers were generally consistent in our infant participants, except in the Late P400 trial set, where pupil dynamics revealed a distinct pattern suggestive of a complex interplay between automatic responses and top–down influences. In these trials, pupils were larger at T1 onset, followed by a stronger constriction than in the Early P400 set, indicating higher attentional engagement or robust attentional refocus on T1 consistent with the deeper processing of the central image observed in that trial set. Due to this particular attentional state, the lateralized T2 may have triggered a rapid pupil response, likely through inhibition of the parasympathetic constriction pathway (40), first step of an orienting response to an unexpected stimulus (Fig. 4C). However, this detection remained implicit as it did not lead to face recognition or overt orienting, as other measures of the T2 perception index (T2_I, Fig. 4E)—notably face-directed saccades—remained unaffected.

We also replicated previous results suggesting a hemispheric asymmetry in face perception in infants (21, 41): interference was shorter for faces presented in the left hemifield, pointing to more efficient right-hemisphere processing. This advantage emerged despite identical low-level features between face and scrambled face stimuli. While this pattern likely reflects face-specific neural mechanisms, we cannot rule out a maturational imbalance in attentional control between the left and right parietal regions that might contribute to the left hemifield advantage. This possibility remains to be further explored.

Finally, we deepened our understanding of the relationship between the N290 and P400 components, traditionally thought to reflect successive stages of face analysis (22). The N290 was not systematically observed after a face image, but was related to its processing depth, as evidenced by responses to T1. Thus, these components should be better described as neural trajectories than obligatory components.

Conclusion

Our study demonstrates the power of combining multiple physiological markers (EEG, eye tracking, pupillometry) and trial-by-trial analysis methods (fly-by, variability quenching) to investigate infant cognition in the absence of overt behavioral reports. It also reveals that even at 4 mo, attention is limited and selective, constrained by the same dynamic bottleneck that characterizes adult’s GWS. This study emphasizes that it is possible to ask complex questions at a young age thanks to brain imaging progress combined with adapted experimental paradigms.

Materials and Methods

Subjects.

19 healthy full-term 4-mo-old infants (mean age 15.4 wk [13.1-17w], 13 girls) were included. This age was chosen to ensure that infants had reached sufficient maturation in both visual sensory perception (42) and attentional control (43, 44) to engage with the complex stimulation paradigm, while still falling within the first semester of life. 23 additional infants were excluded due to insufficient EEG or eye tracking data quality. Collecting high-quality data for both measures proved challenging, and the 10 Hz flicker was probably visually demanding for some participants. The Biomedical Research Ethics Committee (CPP 1000_27, CPP 1000_41) approved the study, and parents provided written informed consent.

Stimuli.

68 colorful images of different teddy bears, and 30 black-and-white images of distinct female faces were selected. An additional 30 scrambled faces were created by rearranging facial features into a bottom-heavy pattern, which infants can distinguish from faces. To control for pupillary reflexes triggered by luminance changes, all stimuli, including the online-generated flicker pattern, were matched for average luminance.

Experimental Paradigm.

Infants sat on their parents’ laps, 65 cm from the screen, in a dark, shielded EEG room. Visual stimuli covered 20° of visual angle, with lateral images 13° from central fixation. The central image appeared on a static gray background and lateral stimuli on a 10 Hz flickering “salt and pepper” background. Eye-tracking (Tobii Tx300) and EEG (128-channel HydroCel Net, Net Amp 300, EGI; 500 Hz, vertex reference) were recorded simultaneously, with EEG sampled at 500 Hz.

Each trial comprised three events (see Fig. 1 for timings). A central teddy bear (T1) was followed by either T2 or nothing (no-T2 trials) at different SOAs. After a fixed T2–T3 interval of 1,600 ms, a random face (T3) appeared on the same side as T2 face, or in no-T2 trials randomly on one side. Trials were pseudorandomized across nine conditions ({no T2, left face, right face} × {SOA 400, 800, 1,200}) and continued as long as infants remained attentive, with brief pauses or termination if refocusing failed.

Analyses.

We combined behavioral and EEG analyses to track infants’ attention and responses to successive visual events. Specifically, overt orientation was measured via saccades toward the lateralized face. For T2, we further computed the percentage of correct gaze orientation to the face vs. the scrambled face, in order to distinguish genuine face recognition from automatic orienting to any visual event. Attentional engagement was indexed by changes in pupil size (39), and covert orienting by SSVEPs. Since both measures are disrupted by eye movements, they provide complementary information to saccade-based metrics. We also computed event-related potentials (ERPs) time-locked to each visual event and notably the temporal expectation dipole (TED), recorded when infants are expecting an event (17, 18). Finally, we analyzed each trial to reconstruct its neural trajectory and assess the depth of attentional engagement with T1.

Eye Tracking Data Analyses.

For each infant, the fixation center was defined by the peak gaze during teddy bear presentation, allowing definition of center, left, and right hemifields. Eye fixations were analyzed using Sam Wass’ infant eye-tracking library (45), and saccades to hemifields were measured after T2 and T3 onsets. Only trials with central fixation at T2 onset and at least 50% of gaze within defined areas were included.

Saccades toward T2 were rare (12 trials/infant; 20%), likely due to the persistent central teddy bear, whereas saccades toward the isolated T3 face were frequent (47 trials/infant; 90%). T2 saccades were categorized as Hits (+1: toward face), False Alarms (–1: opposite side), or No Detection (0). T3 saccade performance was computed by subtracting the proportion of trials with gaze to the incorrect side from those with gaze to the face.

To assess the influence of T2 saccades on T3 behavior, we analyzed trials with saccades at both times using an ordinal LME: T3 saccade outcome (–1: invalid, 0: no saccade, 1: toward face) was predicted by T2 saccade direction (–1, 0, 1), with participant as a random effect.

Pupil Size Analyses.

Trials with saccades between T2 and T3 onset were excluded, as pupil size is unreliable during saccades, yielding an average of 49 trials per infant. Pupil time series were computed using a 150 ms moving average, linear interpolation, and removal of discontinuities. T2 effects on pupil dilation (T2 vs. no-T2 trials) were assessed using cluster-based permutation tests (46).

For the analysis comparing pupil size before T1 onset in Early-P400 and Late-P400 datasets, baseline pupil size was defined earlier, from the 600 ms preceding T1 onset, averaged across three consecutive valid trials. To disentangle an early automatic response to T1 from any potential distinct late effect related to T2, we conducted a post hoc analysis of pupil constriction following T1. The constriction amplitude (minimum after T1 onset) was regressed out of the late dilation response (LMER: Late dilation = β × constriction; β = 0.97, CI: [0.91, 1.03], P < 0.001).

EEG Processing and Analyses.

Continuous EEG recordings were high-pass filtered at 1 Hz and segmented into analysis-dependent epochs (see below). Artifact-contaminated channels were automatically rejected per trial based on amplitude variations. A channel was excluded if its fast average amplitude exceeded 100 μV or if the deviation between fast and slow running averages surpassed 200 μV. Channels marked as bad in over 70% of trials were rejected, as were trials with more than 70% bad channels. Residual bad channels were interpolated using neighboring electrodes. Data were then average-referenced and baseline-corrected.

For T1 and T3, classical visual ERPs were recorded and thus analyzed while no ERP to T2 was seen (see SOM). For T2, we used the neural entrainment by the flickering background to estimate covert orientation to the lateralized stimuli; this analysis was not pertinent for the centered T1, nor for T3 because of too many saccades.

SSVEPs.

They were analyzed on 4.5 s epochs (–1.2 to +3.3 s relative to T2 onset) using the RESS method (47), which enhances signal-to-noise ratio via spatial filtering and allows reconstruction of SSVEP temporal dynamics. The analysis involved two steps: 1) electrode weightings were computed per infant at the first two harmonics (10 Hz, 20 Hz) using peakwidth = 0.7, neighfreq = 1.2, neighwidth = 1.2, sbin = 2, nbin = 11, with maximal weights identified in occipital electrodes at 10 Hz and additional left temporal/parietal weights at 20 Hz; 2) these weightings were applied to reconstruct responses (fwhm = 3 Hz), followed by Hilbert transform. The resulting 1D signals were baseline-corrected (–300 to –100 ms pre-T2), z-scored across harmonics (48), and averaged per infant and condition. Trials with saccades from –200 to +800 ms relative to T2 onset were excluded.

Visual ERP.

Data were low-pass filtered at 15 Hz. T1 epochs ([−500, +4000] ms) were baseline-corrected using the 200 ms pre-T1 interval and averaged by condition (~72 trials per infant, 12 per SOA and T2 conditions). Cluster-based permutation tests identified ERP components typical of infant face responses: P100, N290, and P400 (22).

T3 epochs ([−800, +2200]) ms were baselined on the final 200 ms of T2 display (i.e., -500 to -300 ms relative to T3 onset) to avoid contamination by T2 offset-related responses. As 90% of trials included saccades toward T3 (~500 ± 50 ms latency), analyses focused on early components—P100 (150 to 250 ms) and N290 (300 to 400 ms)—before saccades occurred. Trials were averaged by condition, and mean amplitudes were computed over each cluster within these time windows (21).

Temporal Expectation Dipole (TED).

To identify the infant’s CNV, we followed Mento et Valenza (2016) and applied the mass univariate ERP toolbox (49) across all trials to identify significant activity relative to baseline. Two sustained clusters were observed with topographies consistent with previous reports: an anterior positivity ([700, 2800 ms], Cohen’s d = 1.06, P = 0.001) and a posterior negativity ([650, 2350 ms], Cohen’s d = −1.30, P = 0.002; SI Appendix, Fig. S3A). Dipole strength was estimated by averaging the absolute amplitude of both clusters at each SOA: (SOA400: [600, 1900 ms], t(18) = 3.3, pcorr = 0.006; SOA800: [600, 2300 ms], t(18) = 2.8, pcorr = 0.02; SOA1200: [600, 2700 ms], t(18) = 4.9, pcorr < 0.001, Bonferroni-corrected; SI Appendix, Fig. S3B). T2 effect on TED amplitude was then analyzed by SOA and trial type (Early vs. Late P400).

Flyby Latency to P400 Topography.

Flyby latency refers to the time point in each trial with the topography most similar (based on correlation distance) to a reference template (19). We used the same preprocessed data as above (without baseline correction), normalized across electrodes. The reference P400 common template was the topography at the 450 ms peak of the global field power of the grand-average post-T1. Then we constructed individual templates: For each infant, we identified minima of intertrial variability (VQ events) occurring around 450 ms and selected the one whose topography best correlated with the group-level template. We then sorted trials based on the latency of the closest match (smallest correlation distance) between the topography at each time-point and the individual template. This search was restricted to an 80 ms window centered on the individual’s P400 template. Trials were then split into Early and Late P400 sets using the median latency per SOA and face side. We confirmed no main effect of T2 or category on trial count (T2: F(1,18) = 1.1, P = 0.3; Category: F(1,18) < 1).

VQ.

This measure captures the similarity of neural trajectories across trials at each time point. It was computed by estimating correlation distances across trials within each dataset. Significant time windows were estimated using cluster-based permutation tests within the first second after T1 onset.

T2 Statistical Analyses.

The design included six conditions: SOA (3 levels) × T2 (2 levels) with face presentation balanced across hemifields given prior evidence of asymmetrical face processing in infancy (21). However, trial numbers were unbalanced across conditions due to two experimental features: 1) To reduce overall experiment duration, no-T2 trials at 400 and 800 ms SOAs were set to 10% of trials in pseudorandom blocks compared to 20% in the other four conditions. To maintain comparability, no-T2 trials at 1,200 ms SOA were randomly subsampled to match shorter SOA counts. 2) Trials were excluded when infants’ gaze was not properly centered at T2 onset—a more frequent issue at short SOAs due to delayed orientation to the central teddy bear.

To analyze attentional responses to T2—pupil size, gaze behavior, SSVEPs, and hemispheric asymmetry—we used linear and generalized linear mixed-effects models (LME, GLME; binomial family for saccade rates) from the lme4, lmerTest, and ordinal R packages. Normality was assessed via QQ plots, and model comparisons were performed using Akaike’s Information Criterion (AIC) to test additive vs. interaction effects. All statistical tests were two-tailed unless otherwise specified (i.e., face orientation analyses, for which directional hypotheses were defined a priori). Multiple comparisons were corrected where appropriate, using Bonferroni correction for post hoc comparisons across SOAs in the absence of significant interactions and cluster-based permutation tests for time-series analyses (46). An exception was the saccade data toward T2, where no correction was applied due to limited trial numbers; this analysis primarily served to validate pupil and SSVEPs findings.

To address the power limitations of independent analyses, we constructed a composite T2 perception index (T2_I) aggregating all relevant measures: pupil dilation, SSVEP reduction, TED amplitude, saccade rate, and saccade performance (SI Appendix). Independence between saccade rate and saccade performance was confirmed (Kendall’s tau = 0.13, z = 1, P = 0.3). Each measure was z-scored, and T2 vs. no-T2 differences were computed per participant.

T3 Statistical Analyses.

As in T2 analyses, trial numbers were unbalanced: the no-T2 condition had 30% fewer trials than T2 condition; SOA 400 and 800 ms had 40% and 30% fewer trials, respectively, than SOA 1,200 ms; and the Low-T2 set had 30% fewer trials than the High-T2 set. In the ERP analysis, trials were pooled across hemifields for each SOA after confirming that T3 presentation side had no effect in the no-T2 condition at SOAs 400 or 800 ms. After pooling, trial counts no longer differed significantly across conditions (GLME, Poisson family; additive model: SOA 800, P = 1.0; SOA 1200, P = 0.12; T2, P = 0.7). Normality was assessed using the Shapiro test, and standard ANOVAs were then applied to examine SOA and T2 effects.

For saccade performance, which was T3-side dependent, we used an ordinal mixed-effects model with three outcomes: hit (face), no detection, and false alarm (opposite side), applied to test both SOA and T2 perception effects.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The individual-level infant data used in this study involve sensitive information collected from human participants, with parental consent permitting use strictly for scientific research and prohibiting commercial applications. Due to privacy and ethical restrictions, these data cannot be made publicly available. However, qualified researchers affiliated with public research institutions may request access by outlining their scientific objectives, including replication studies, by contacting the corresponding author. Requests will be reviewed by the ethical officer of the Commissariat à l’Énergie Atomique et aux Énergies Alternatives (CEA), the promoting institution.