Effects of Statistical Learning on Stimulus Detection Speed Reflect Attention Rather Than Representational Changes

Abstract

Statistical learning (SL) is believed to enable humans to assimilate a range of statistical structures, and thus plays a role in many cognitive functions. There is also a growing interest in how SL interacts with basic cognitive processes, including perception and attention. Here, we ask how the extent to which stimuli predict, and are predicted by, other elements in a continuous stream affects their perception (i.e., encoding and representation) and the attention they attract. In two experiments, participants were first exposed to a stream of structured pairs of visual shapes (e.g., AB and CD). Then, they completed a target detection task to test if stimulus detection speed is influenced by a target's predictability during exposure (indexing representation) or by whether the shape preceding the target reliably predicted elements in the input (indexing attention). In Experiment 1 (N = 86), Reaction Times (RTs) were faster for second elements from structured pairs (i.e., cued elements) than first elements (i.e., cue elements), even when they appeared in new configurations (e.g., AD and CB). However, in Experiment 2 (N = 89), which orthogonally manipulated the target and the preceding shape properties, RTs were influenced solely by whether the preceding element was a cue for other elements, but not by the target's predictability. Thus, in contrast to previous studies using the same paradigm, our results do not provide evidence for an effect of SL on representation. Instead, our findings highlight how attention is guided by knowledge of statistical regularities, pointing to SL as a system that helps minimize uncertainty in structured environments.

1. Introduction

The seminal study by Saffran, Aslin, and Newport (1996), which revealed that infants can parse continuous linguistic input based on the co‐occurrence statistics of its elements, has inspired hundreds of studies that examined how humans extract the statistical regularities of different types of sensory inputs across space and time. Indeed, Statistical learning (SL) has since been demonstrated in humans in the visual (e.g., Fiser & Aslin, 2001), verbal auditory (e.g., Saffran, Johnson, Aslin, & Newport, 1999), nonverbal auditory (e.g., Creel, Newport, & Aslin, 2004), and tactile domains (e.g., Conway & Christiansen, 2005), across various contexts. Given its ubiquitous nature, it is now believed that SL plays a fundamental role in various low and high cognitive functions: Historically, much of the interest in SL came from its presumed role in language learning and processing (see Romberg & Saffran, 2010; Siegelman, 2020), but more recently, it has been suggested to play a role also in other abilities, such as motor learning, social skills, and music abilities (see, e.g., Kushnir, Xu, & Wellman, 2010; Monroy, Meyer, Schröer, Gerson, & Hunnius, 2019).

Notwithstanding the continuously accumulating empirical data on SL, a key challenge remains in characterizing the interactions between SL and other basic cognitive mechanisms (see Frost, Armstrong, & Christiansen, 2019, for a detailed discussion). This is despite an emerging literature examining these interactions (reviewed below), and the fact that they are of clear theoretical interest: Given that SL helps individuals assimilate the structure of their environment, and given that human surroundings are full of structure, SL most likely modulates multiple basic cognitive processes. Such interactions are also in line with predictive processing theories, which posit that the brain actively anticipates incoming input, with sensory information projected onto a predictive model of the environment (Clark, 2013; Friston, 2010; Friston, Kilner, & Harrison, 2006). The focus of the current work is on SL's interactions with attention and perception: That is, we ask how learning of statistical structure may modulate what individuals attend to in a stream of information, and the way in which individuals perceive the elements that comprise that stream of information (i.e., how they encode elements to form stable mental representations of them).

In terms of attention, there has indeed been a growing interest in examining how sensitivity to statistical structure impacts attention allocation. For example, Kidd, Piantadosi, and Aslin (2012, 2014) demonstrated what they termed “The Goldilocks Effect,” where infants’ attention is preferentially drawn to events of moderate levels of predictability, rather than to overly predictable or unpredictable ones. Forest, Siegelman, and Finn (2022) further showed that adults gradually shift their attention from simple stimulus streams to more complex ones. Moreover, in spatial SL paradigms, it was found that participants’ gaze follows contingencies by going between two shapes that co‐occur often (Arató, Rothkopf, & Fiser, 2024), and that SL helps direct participants’ attention toward probable target locations (Theeuwes, Bogaerts, & van Moorselaar, 2022) and away from common distractor locations (Wang & Theeuwes, 2018a, 2018b; Duncan et al., 2026). These findings also cohere with the conceptualization of attention as a “precision gate” to predictive processing (Feldman & Friston, 2010; Smout, Tang, Garrido, & Mattingley, 2019). In this view, the predictive model, which is shaped by prior statistical knowledge, guides where attention is deployed.

In terms of the interactions between SL and perception, SL models posit that the constraints of the sensory systems (responsible for the encoding of elements into representations) place limits on what statistical regularities can be subsequently learned (e.g., Frost, Armstrong, Siegelman, & Christiansen, 2015). These constraints are thought to then lead to modality‐ and domain‐specific effects observed in SL tasks (e.g., Bogaerts, Siegelman, Christiansen, & Frost, 2022; Emberson, Conway, & Christiansen, 2011). In parallel, there are reasons to believe that sensitivity to co‐occurrence information shapes the encoding and the resulting representations of the individual stimuli that form these structures. Predictive processing theories suggest that sensory input is projected onto prior statistical knowledge that is encoded within sensory systems (Friston et al., 2006; Friston, 2010). Consistent with this view, it was found that statistical structure in a sensory stream impacts activation already in early sensory cortices (Nastase, Iacovella, & Hasson, 2014; Tremblay et al., 2013). Moreover, neuroimaging SL studies have shown that stimuli that co‐occur have more similar representations (Schapiro, Kustner, & Turk‐Browne, 2012, 2016).

Behaviorally, however, only a few studies have examined how the encoding and representation of individual elements are shaped by the input's statistical structure. One such study is the work by Barakat, Seitz, and Shams (2013) on visual and auditory SL, which examined how exposure to a structured input affects the speed of detection of different individual elements from the stream, taken as a proxy for the robustness of the stimulus representations (under the assumption that more robust representations should lead to faster detection speed). In this study, participants were first exposed to a standard familiarization phase, where six fully regular stimulus pairs (i.e., with a transitional probability, TP, of 1) appeared in a pseudorandom order. Then, following exposure, participants completed a target detection task, where, in each trial, they had to detect one target in a stimulus stream. The target was either a predictable element from the exposure phase (e.g., element B from the AB pair, hereinafter: a cued element), or an unpredictable stimulus that predicted the following stimulus (e.g., A, hereinafter: a cue element). Crucially, whereas in half the trials the target appeared as part of its structured pair, in other trials, the target “switched” its pair membership, appearing with a stimulus from another pair (e.g., given AB and CD, the target D appeared after A, rather than C). Barakat et al. (2013) documented an expected facilitation to cued versus cue stimuli in their learned context (as typically found in target detection tasks, e.g., Batterink & Paller, 2017; Fiser & Aslin, 2002; Turk‐Browne, Jungé, & Scholl, 2005). Interestingly, they also found that cued stimuli were recognized faster than cue stimuli outside their learned context (e.g., facilitation for D even when it appeared after A). These findings were taken to reflect representational changes: Thus, per the authors’ interpretation, cued (i.e., predictable) stimuli have more robust representations than cue (i.e., unpredictable) elements, leading to their faster recognition. Barakat et al. (2013) also proposed potential mechanisms underlying this effect: increased attention allocated to cued stimuli during exposure (i.e., participants anticipate the appearance of cued stimuli during learning, which leads to better encoding and more robust representations), or greater internal rewards by cued stimuli, which again strengthen their representations.

While highly intriguing, this previous study had two important limitations. The first is statistical power: The conditions showing facilitation outside their learned context had small sample sizes (i.e., N = 12 and N = 31 for the visual and auditory conditions, respectively), and thus a larger sample size is needed to establish the findings’ robustness. Second, and importantly, Barakat et al.’s design does not disentangle the effect of attention from that of the stability of the representations. Thus, in their design, cued elements appeared with some cue elements during the target detection phase (i.e., D appearing after A, which by itself was a cue of another shape). Therefore, a facilitation for D could be either due to (1) a more precise representation of cued elements (i.e., a representation account), but also (2) more attention allocated to elements that appear following cue elements in the target detection phase (i.e., an attention account). In other words, if following learning, learners allocate their attention to elements that follow a cue element (regardless of the target's status), then Barakat et al.’s findings could be explained by how attention is allocated during the target detection phase, rather than representational changes.

The goal of the current study is to address both limitations. In Experiment 1, we report a high‐powered replication of Barakat et al. (2013) to establish the reproducibility of the original findings. Then, in Experiment 2, we use a modified design, where we orthogonally manipulate the target's status, and whether the element preceding it typically predicted other elements during exposure. This modified design allows us to dissociate representational processes from effects related to attention. To preview our findings, we show that whereas the results by Barakat et al. replicate fully, there is no evidence that the observed effect stems from representational changes. Rather, we find that following learning, attention is allocated specifically to elements that appear after other cue stimuli. These findings highlight how SL interacts with attentional processes, potentially boosting learners’ attention toward places that are expected to be lower in uncertainty, an interpretation we return to in the General Discussion.

2. Experiment 1

2.1. Methods

2.1.1. Participants

A sample of N = 100 English‐speaking participants, recruited via Prolific, participated in the experiment. Participation was limited to participants from the UK and the US, aged 18–40. The sample size was determined using a priori power analysis, assuming the statistical procedure used by Barakat et al. (a paired sample two‐tailed t‐test), a minimal effect size of d = 0.3, and statistical power of 0.8, while accounting for 10% attrition. The minimal effect size used was conservatively smaller than the one in Barakat (d = 0.65 for the crucial comparison). Both Experiments in this paper were conducted in accordance with the Declaration of Helsinki and were approved by the institution's ethics review board (IRB_2024_019).

2.1.2. Stimuli and procedure

The experiment utilized 16 visual shapes from previous SL studies (e.g., Siegelman, Bogaerts, Kronenfeld, & Frost, 2018). Of these 16 shapes, 10 lists were randomly assembled, each with 12 shapes to serve as stimuli during the familiarization phase, with the remaining four used as foils in the shape recognition test (see below). Lists were randomly assigned to participants.

The experiment comprised three phases (Fig. 1, Panels a−c). The first was a typical visual SL familiarization phase, consisting of six fully regular pairs (within‐pair TP of 1). The six pairs created a pseudorandomized stream, where each pair appeared 75 times, with the only constraint that the same pair could not appear twice in a row (i.e., between‐pair TP of 0.2). We note that 75 repetitions exceed what is typically used in similar SL tasks and should result in participants’ learning reaching a plateau (e.g., Siegelman et al., 2018). We have opted to use this number to maintain similarity with Barakat et al. (2013), which used an even higher number of 150 repetitions, without overtaxing our participants. Shapes were presented in the middle of the screen one after the other, for 300 ms each, with an inter‐stimulus interval (ISI) of 100 ms. Before familiarization, participants were told that they would be presented with a stream of shapes for about 6 min, and were instructed to pay attention as they would later be tested. Participants were not informed about the fact that shapes co‐occurred, or about the nature of the following test (full task instructions are available via the project's OSF page).

Fig. 1.

The design of Experiments 1 and 2. Panel (a): Twelve shapes create six pairs. Panel (b): Participants observe a pseudorandomized exposure stream (75 repetitions per pair). Panel (c): Target detection phase, Experiment 1: Targets are either cue (the first element in a pair) or cued (the second), appear with either a match (complementary shape) or mismatch (a shape from a different pair) element. Panel (d): Target detection phase, Experiment 2: Targets are cue or cued, appearing after a cue shape or after a cued shape.

Following the familiarization phase, a shape recognition test began, used as a screening task to exclude inattentive participants. In this task, participants completed 16 2‐alternative‐forced‐choice (2‐AFC) trials, each presenting a single randomly selected shape from the 16‐shapes pool. Participants were asked to indicate whether each shape appeared in the first part of the study or not. We decided a priori to exclude participants who would not reach a critical threshold of 12/16 correct responses (reflecting significant above‐chance performance according to a binomial distribution; see also Bogaerts, Siegelman, & Frost, 2016; Siegelman et al., 2019).

The third phase of the experiment was the central target detection phase. Each trial began with a shape presented in the middle of the screen for 3 s. Participants were told that this shape was their target for that trial and were instructed to press the spacebar as quickly as possible when it appeared in the following shape sequence. Then, participants were presented with a sequence of 10 shapes, one of which was the target. This task consisted of 120 trials, with each of the 12 shapes from the familiarization phase serving as the target 10 times. In each sequence, 8 out of 10 shapes were made up of four structured shape pairs (i.e., the structure of the familiarization phase was maintained), with the target appearing either with or without its paired shape (see below). The target always appeared between the third and the eighth position in the trial. As in the familiarization phase, each shape appeared for 300 ms with a 100 ms ISI.

Importantly, the target detection task had four within‐subject conditions, in a 2 (Target type: cue vs. cued) × 2 (Trial type: match vs. mismatch) design. First, in half the trials, the target was a cue target (i.e., the first element in a pair from the familiarization phase), whereas in the other half of trials, the target was a cued type (i.e., the second element from a pair). Second, in half of the trials, the shape paired with the target was replaced by a corresponding shape from the sixth pair from the familiarization phase (i.e., the one not used in this trial; a mismatch trial), whereas in other trials, the target was presented with its originally paired shape (a match trial). Overall, then, the target could be either cue (i.e., unpredictable during familiarization) or cued (i.e., predictable during familiarization), appearing either with its originally paired shape (i.e., match) or with another shape (i.e., mismatch). Note that in the case of cue targets, the manipulation of Trial Type was only apparent after the target's appearance (i.e., match or unmatched stimuli only appeared after the cue stimuli). This is why the main effect of Trial type in the models below is not of theoretical interest.

2.2. Results

2.2.1. Data preprocessing

We removed the data of N = 7 participants who failed to achieve the threshold in the shape recognition task. We further removed: trials in the target detection task in which participants pressed the spacebar before the target appeared (10.5% of the trials); the data of N = 3 participants with more than 40% of trials with early press; all trials in which participants failed to respond (7.5% of the trials); the data of N = 4 participants with more than 40% of the target detection trials with no response,¹ and trials with Reaction Times (RTs) below 250 ms (2% of trials). The final sample comprised N = 86 participants, with 100.2 trials per participant on average. For these remaining participants, accuracy in the shape recognition task was expectedly high, M = 92.5% (SD = 7.9%).

To ensure that our central findings hold under different and additional outlier removal procedures (e.g., more lenient lower RT thresholds enabling faster responses to be included; additional higher RT thresholds excluding slow responses), we conducted a robustness analysis reported in Supplementary Materials 1. We note that our central findings in both Experiments are robust across the vast majority of preprocessing parameters tested (see Supplementary Materials 1).

2.2.2. The effect of Target and Trial Type on target detection

We used a linear mixed‐effect model,² with trial‐level RT after natural log‐transformation as the dependent variable,³ and fixed‐effects for Trial Type (match vs. mismatch; effect‐coded), Target Type (cue vs. cued; effect‐coded), and their interaction, with additional fixed‐effects for the target's index within the stream (3−8, numeric) and trial number (1−120), included as controls. The model also included by‐participant and by‐item (i.e., target) random intercepts, as well as by‐participant random slopes for Target Type (the maximal random effect structure that converged; Barr, Levy, Scheepers, & Tily, 2013). Models were fitted using the lme4 package in R (Bates, Mächler, Bolker, & Walker, 2015), with significance estimated via lmerTest (Kuznetsova, Brockhoff, & Christensen, 2017). Fixed‐effect estimates are presented in Table 1.

Table 1.

Fixed‐effects estimates in a model predicting log‐transformed RTs in the target detection task, Experiment 1

Predictor	Estimate	SE	t	p
Intercept	6.15	0.019	314.44	<.0001
Target type: cued	−0.008	0.003	−2.39	.018
Trial type: mismatch	0.004	0.002	1.8	.070
Target type × Trial type	−0.0007	0.002	−0.29	.764
Target index	−0.029	0.001	−18.78	<.0001
Trial number	0.0005	<0.001	7.85	<.0001

Replicating Barakat et al. (2013), we found a significant effect for Target Type on log‐transformed RT (β = −0.008, SD = 0.003, p = .018), with cued (i.e., second in pairs) shapes recognized faster than cue (i.e., first in pairs) shapes. Also, in line with Barakat et al. (2013), there was no significant effect either for Trial Type (β = 0.004, SD = 0.002, p = .07) or, importantly, for the interaction between Trial Type and Target Type (β = −0.0007, SD = 0.002, p = .764; see Fig. 2). Indeed, post‐hoc contrasts revealed similar effects of Target Type within the two levels of Trial Type (in match trials: Ψ = 0.016, SD = 0.008, 95% CI [−0.001, 0.033]; in mismatch trials Ψ = 0.019, SD = 0.008, 95% CI [0.001, 0.036]). Overall, these results indicate that cued (i.e., predictable) shapes were consistently recognized faster than cue (i.e., unpredictable) shapes, even when paired with unmatched stimuli.

Fig. 2.

Estimated effects of target type (cue vs. cued) and trial type (match vs. mismatch) on RTs in Experiment 1. Error bars represent standard errors. For interpretability, log‐transformed values transformed back to raw RT.

3. Experiment 2

Experiment 1 showed that target detection is influenced by the statistical structure of an exposure stream in which stimuli were first presented: Specifically, targets that were cued (i.e., predictable) during the exposure phase were recognized faster than cue (i.e., unpredictable) targets, even when appearing after shapes that served as a cue of other, nontarget, elements (i.e., mismatch trials). This provides a conceptual replication of Barakat et al. (2013).

However, as mentioned above, such facilitation is consistent with two theoretical explanations. The first is that SL may impact stimulus representation, with cued (i.e., predictable) elements having more robust representations because they were encoded better. This explanation entails that cued targets would be identified faster regardless of where they appear during the target detection stream. The second has to do with attention allocation: Due to SL, attention may shift during the target detection phase to elements that follow (i.e., appear after) a cue element, that is, an element that predicted other elements during exposure. This explanation entails that any shape would be recognized faster when it appears after a cue shape, not only cued shapes. Importantly, the design of Experiment 1 (and Barakat et al., 2013) cannot disentangle these explanations: Cued elements always appeared after a cue element in the target detection phase, even when it was not the element specifically matched with them. The goal of Experiment 2, therefore, is to orthogonally manipulate the target's properties and the properties of the element preceding it.

3.1. Methods

3.1.1. Participants, stimuli, and procedure

We recruited N = 100 additional English‐speaking participants from Prolific, who did not take part in Experiment 1. The exposure and shape recognition phases were identical to Experiment 1. The target detection task again had a 2×2 within‐participant design. However, whereas in Experiment 1, cued targets always appeared after cue elements and vice versa, in Experiment 2, we manipulated the type of the shape before the target (see Fig. 1d). Thus, in half the trials, the target appeared after a cue‐type shape, and in half of trials, after a cued‐type shape (we refer to this two‐level factor below as Trial Type: After‐cue and after‐cued). The other factor was again the Target Type: In half of the trials, the target was of cue type, and in the other half, it was of cued type. Note that in Experiment 2, targets did not appear with the shapes from the corresponding learned pairs, and were, therefore, not locally predictable during the target detection phase (however, all eight other stimuli in each target detection trial did appear in their ordered pairs, meaning that the learned structure was generally maintained during this phase).

3.2. Results

3.2.1. Data preprocessing

We followed the same data cleaning procedures of Experiment 1: We removed N = 2 participants who did not complete the task; N = 6 participants who failed to reach the shape recognition threshold; trials with early presses (10.5% of trials) and N = 2 participants with more than 40% of early presses; trials with no response (2.5% of trials); and trials with responses shorter than 250 ms (9.4% of trials; but again see Supplementary Materials 1 for robustness analysis). The final sample comprised N = 89 participants with 103.9 trials per participant on average. Participants’ shape recognition accuracy was on par with Experiment 1 (M = 93.2%, SD = 6.5%).

3.2.2. The effect of Target and Trial Type

We employed the same linear mixed‐effect model as in Experiment 1, with by‐participant and by‐target random intercepts and by‐participant random slopes for Target Type (Table 2 shows fixed‐effect estimates). We found a significant effect of Trial Type on log‐transformed RT (β = 0.007, SD = 0.003, p = .013) with targets appearing after cue shapes recognized faster than targets appearing after cued shapes. Importantly, there was no significant effect of Target Type (β = −0.0007, SD = 0.003, p = .841): That is, no evidence for an RT difference between cue and cued targets. Furthermore, there was no evidence for an interaction between Target Type and Trial Type (β = −0.001, SD = 0.002, p = .554; Fig. 3), with post‐hoc contrasts again revealing similar effects of Target Type within the two levels of Trial Type (in after‐cue trials: Ψ = −0.001, SD = 0.008, 95% CI [−0.018, 0.015]; in after‐cued trials Ψ = 0.004, SD = 0.008, 95% CI [−0.012, 0.021]). Overall, these results indicate that any target appearing after a cue shape was recognized faster, regardless of its own predictability during the exposure phase.

Table 2.

Fixed‐effects estimates in a model predicting log‐transformed RTs in the target detection task, Experiment 2

Predictor	Estimate	SE	t	p
Intercept	6.16	0.015	390.49	<.0001
Target type: cued	−0.0007	0.003	−0.20	.841
Trial type: after cued	0.007	0.003	2.52	.013
Target type × Trial type	−0.001	0.002	−0.59	.554
Target index	−0.031	0.001	−20.57	<.0001
Trial number	0.0006	<0.001	8.75	<.0001

Fig. 3.

Estimated effects of trial type (after‐cue vs. after‐cued) and target type (cue vs. cued) on RTs in Experiment 2. Error bars represent standard errors. For interpretability, log‐transformed values transformed back to raw RT.

4. General discussion

The current study was set out to examine what determines shape identification speed following an exposure to a statistically structured visual input, as a means to investigate the effects of SL on stimulus representation and on attention. Our findings point to a clear conclusion: Shape detection was facilitated not due to a target predictability during exposure (cf. Barakat et al., 2013) but rather by whether the element preceding it predicted other elements during exposure. Our findings, therefore, suggest that it is attention mechanisms that are in play in determining shape recognition times, by guiding attentional resources toward events that are expected to be highly predictable by the preceding element—that is, places of low uncertainty in the stream. In contrast, in a design that orthogonally manipulated targets’ predictability and the preceding shape properties, we found no evidence for representational changes following visual SL (cf. Barakat et al., 2013).

At a theoretical level, our findings join recent studies in unveiling how attention allocation is shaped by the environment's statistical structure. Our findings underscore the bidirectional relation between SL and attention: What we attend affects what we learn (e.g., Arató et al., 2024; Duncan & Theeuwes., 2020), and what we learn impacts what we attend (e.g., Forest et al., 2022; Kidd et al., 2012; Duncan et al., 2026). More broadly, our findings emphasize the interactions between SL and other basic cognitive processes, and its integrated nature in basic functions of cognition.

Our main finding, then, is RT facilitation to elements that appear following a cue element (i.e., one that predicted other elements during exposure). What is still unclear, however, is the exact conditions under which such facilitation happens. Specifically, there are two potential driving forces that can lead to our findings. The first is that attention is drawn toward familiar events (see, e.g., Lancry‐Dayan, Nahari, Ben‐Shakhar, & Pertzov, 2018). That is, when participants develop knowledge of a statistical pattern (e.g., AB), their attention may be channeled toward what they anticipate knowing, even before an event appears (e.g., what comes after A; Zhao, Al‐Aidroos, & Turk‐Browne, 2013). In other words, learners may channel their attention toward what they expect to be a familiar event (even when it is ultimately surprising). A second explanation, in contrast, is that attention prioritizes informative input, a suggestion in line with recent claims about the role of SL in information foraging (Frost et al., 2025). Thus, according to this framework, SL serves as a tool for detecting novel covariates in dynamic environments. From this perspective, the observed facilitation we see may be related to the fact that during the target detection phase, the learned structural regularities shifted from what was learned during exposure (i.e., due to the presentation of statistical incoherent information). These deviations may have made the events following cue elements particularly informative (see also Smout et al., 2019; Stilwell, Bahle, & Vecera, 2019). Put differently, whereas the first account suggests a negative correlation between entropy and attention (i.e., more familiarity leads to greater attention), the second account may predict a nonlinear and time‐dependent relation (see evidence by Kidd et al., 2012, Forest et al., 2022, in a different empirical context). Although we cannot disentangle these two explanations with the current data, they could be empirically tested in future studies by examining the relation between attention and uncertainty (i.e., entropy) more continuously.

Relatedly, another question in regard to our findings is whether they also reflect attentional competition. Specifically, if predictable elements draw more attention to become more salient (see Duncan et al., 2026), responses following cued shapes might be slowed due to a temporary depletion or “bottleneck” of attentional resources (Raymond, Shapiro, & Arnell, 1992; Zivony & Lamy, 2022). From this perspective, it is possible that attentional competition contributed to the RT difference we observed for targets appearing after‐cue and after‐cued shapes in Experiment 2. However, we contend that an account of attentional competition is incomplete: If one assumes prioritization of predictable elements, they will arguably also expect general facilitation to the cued elements themselves, a result that was not observed (see null effect for Target Type in Experiment 2). Overall, then, whereas attentional competition following cued elements is a relevant contributing factor in explaining the impact of SL on attention in other experimental contexts (see Clement & Anderson, 2023; Gao & Theeuwes, 2020), a more parsimonious account of the current findings seems to be one related to facilitation following cue elements.

Importantly, our findings also provide no evidence for an effect of SL on representational processes. Our results thus refine Barakat et al.’s (2013) conclusions, highlighting that at least in the current settings, it is attention, rather than representational changes, that leads to facilitation in item recognition. Still, our findings do not preclude the possibility that SL can lead to representational changes. Indeed, using a different task (i.e., a masking procedure with a 2‐AFC task), Barakat et al. (2013) did find evidence for such an SL−representation interaction (see also Umemoto, Scolari, Vogel, & Awh, 2010), and neuroimaging studies pointed to modulation of representations following exposure to statistical regularities (Schapiro et al., 2012; 2016). Moreover, even if fully predictable stimuli (i.e., TP of 1) do not undergo representational changes, other statistical regularities may induce such modifications. In particular, under predictive processing perspectives, stimuli that consistently lead to violation of expectations may be better represented in comparison to neutral stimuli, because they incorporate more prediction error (i.e., learning signal), resulting in representational resources devoted specifically to them (see Greco, Moser, Preissl, & Siegel, 2024). Future experiments should explore whether and how representational changes manifest across different experimental settings, varying both the regularities and tasks employed.

Here, it is also important to discuss a surprising null effect we observed, specifically, the fact that in our paradigm, the detection of cued targets was not slowed down by presenting them outside their learned context (i.e., in mismatch trials). This result suggests that the prediction of specific elements does not lead to their faster recognition. We first would like to consider this finding in a broader empirical context, taking into account findings from other experimental paradigms. In particular, studies using the alternating serial reaction time (ASRT) task do provide evidence for facilitation of predictable versus unpredictable stimuli (i.e., stimulus‐specific predictions; e.g., Howard & Howard, 1997; Song, Howard, & Howard, 2008). This seeming discrepancy, however, should be interpreted given the differences between the paradigms, which include the nature of the learned regularities (i.e., whether they are deterministic or probabilistic), how knowledge is probed (i.e., via an offline target detection phase, or an online measurement throughout exposure), and whether and to what extent the learned structure is violated (recall that in our variant of the target detection task, in half the trials, there is violation of structure). These differences may then invoke different processes. For instance, in a task such as ASRT, where there is little structure violation during testing, participants may inhibit a response following the appearance of an unpredictable stimulus, as they are predicting a different stimulus. In contrast, in tasks where many test trials violate expectations, such as our design, this response inhibition to unpredictable elements may be reduced (see also Bugg, Jacoby, & Toth, 2008; Rothermund, Gollnick, & Giesen, 2022). Instead, in such designs, participants may more readily respond to any stimulus appearing after a cue element (i.e., elevated responsiveness elicited independently of the identity of the target). These effects may further vary depending on the complexity of the regularities, which can subsequently also impact participants’ awareness of the structure (i.e., more stimulus‐specific predictions when learning is more implicit). These stimulus‐specific and stimulus‐independent mechanisms should be further studied in future work.

From this perspective, it is also interesting to re‐evaluate claims from other SL studies using target detection paradigms (e.g., Bertels, Boursain, Destrebecqz, & Gaillard, 2015; Franco, Gaillard, Cleeremans, & Destrebecqz, 2015; Lukics and Lukács, 2021). Generally, the target detection task has often been seen as reflecting prediction, given a learned pattern (e.g., AB), of a cued element (B) from a cue element (A). However, as noted above, our results point to a different interpretation: It is not the prediction of a specific shape that leads to facilitation in this paradigm, but rather a shift in attention given the cue shape before a target. In other words, the facilitation observed in typical target detection tasks is not a pure measure of the prediction of a specific stimulus. This is important to consider when planning future studies: Performance in the target detection task may not reflect the learning of specific regularities, but rather how attentional resources are allocated to different parts of the input. As such, this task should be used with caution in studies that require a task not impacted by attentional processes.

Finally, we note that all our theoretical conclusions are based on the analysis of response latencies (i.e., RT being the dependent variable), with an attempt to manipulate design properties to disentangle between processes broadly related to attention and representational changes. Clearly, however, RT is a coarse‐grained composite measure, which is influenced by several processes, such as motor preparation, expectations, and changes to decision criteria, all relevant to these broad cognitive functions. We leave it for future work to isolate more specific mechanisms involved in the effects we observe, using, for example, more targeted computational models (e.g., Pedersen & Frank, 2020) as well as Electroencephalography (EEG) measures (e.g., Bogaerts, Richter, Landau, & Frost, 2020).

5. Concluding remarks

Overall, our study highlights the dynamic interplay between SL and attentional processes. By demonstrating that attention is guided toward low‐uncertainty regions of a stimulus stream, we illuminate that learners change how they attend to different parts of the input given its statistical structure, and, more broadly, the importance of investigating the interactions between SL and basic cognitive functions. Future research should continue to explore these relations, examining how SL modulates attention, perception, and other basic cognitive processes across diverse experimental contexts.

Conflicts of interest

The authors declare no conflicts of interest.

Ethics approval statement

Behavioral experiments have been performed in compliance with relevant laws and institutional guidelines regarding human participants and have been approved by the institution's ethics committee (approval #: IRB_2024_019). Informed consent was obtained from all participants.

Supporting information

Supporting Information

Data availability statement

The raw data, analysis script, and experimental software code and program are available at: https://osf.io/ps4af/.