Prior to the reorganization of diagnostic criteria in the DSM-5 (American Psychiatric Association, 2013), ‘dyslexia’ meant an array of impediments characterized by poor word recognition, decoding, and spelling abilities (i.e., difficulty in ‘learning to read’). Additionally, specific subtypes of dyslexia were also thought to exist, depending on severity of the disability. With little consensus over these complex subtypes, the disorder eventually became housed under a broader classification that focused more on both the heterogeneity within dyslexia as well as its co-occurrence with other disorders (Snowling et al., 2020). Currently, developmental dyslexia is classified under ‘developmental disorders’ (Ullman et al., 2020); and to be included within this category an individual must experience difficulties in various aspects of language learning (Bishop, 2017) such as learning to read and spell. Snowling et al. (2020) argue that dyslexia should be considered as less of a purely reading disorder and more of a persistent difficulty with decoding and spelling fluency, markedly affecting academic performance. Thus, rather than referring to a specific reading disorder, the classification of ‘dyslexia’ henceforth reflects these updates. Dyslexia is often comorbid with other cognitive disabilities such as arithmetic learning disability (von Aster and Shalev, 2007; Dirks et al., 2008; Landerl and Moll, 2010), attention deficit hyperactivity disorder (Gilger et al., 1992; Pennington, 2005), specific speech disorder (Pennington and Bishop, 2009), developmental language disorders (Catts et al., 2005; Snowling et al., 2020) and even developmental coordination disorder (Kaplan et al., 1998). This suggests that reading ability is heavily intertwined with other aspects of cognition and academic skills required for typical development as has been recently outlined by Moll et al. (2020). These patterns of comorbidity raise the possibility that at least for some individuals, the problems are due to more general issues with attention, learning, or other cognitive processes that span across these different domains. Different patterns of comorbidities may complicate the interpretation of different studies that assess statistical learning in individuals with dyslexia.

To better understand how and why statistical learning deficits might underlie developmental dyslexia next, we review prominent deficit-centered theories of dyslexia and their relevance to statistical learning. By reading through the framework below it is apparent that some of these theories are consistent with the notion of more general and pervasive cognitive deficits. Figure 1 summarizes these theories and is organized into four categories based on the postulated type of deficit: sounds and phonemes, cognitive, neurobiological, and multiple deficit theories.

First, considering theories related to sounds and phonemes (Figure 1A), the phonological deficit theory is perhaps the most prominent perspective (Vellutino et al., 2004). Here decoding involves analyzing printed symbol forms (e.g., letters) and mapping them to the sounds to which they refer (e.g., that the word ‘pot’ begins with a /p/ sound; Castles et al., 2018). Under this view, reading impairment results from a disruption in the grapheme-phoneme mapping learned during typical reading acquisition. However, in recent years, the existence of a purely phonological deficit has been challenged (Banai and Ahissar, 2018; Cabbage et al., 2018; Lachmann, 2018; Giofrè et al., 2019). Instead, other sound-related theories such as the amplitude modulation rise-time deficit theory (Goswami et al., 2002) have been proposed.

If dyslexia is at its core a problem with processing sounds and phonemes, as these two theories suggest, then we might expect a bigger impairment in dyslexia for statistical learning tasks using auditory and phonological-based stimuli compared to non-auditory or non-phonological stimuli. Accordingly, statistical learning tasks involving visual or visual-motor patterns would be expected to be less affected. If such a result was found, then it would be difficult to conclude that dyslexia is associated with domain-general statistical learning impairments but rather is a sound or phoneme-based disorder as this group of theories imply.

Some other theories focus on more general cognitive processes (Figure 1B). These theories propose impairments in visual attention (Bosse et al., 2007), serial order learning (Szmalec et al., 2011), temporal processing and sampling (Tallal, 1984; Goswami, 2011), memory, and the ability to track and/or disengage from environmental stimuli (Sperling et al., 2005; Ahissar, 2007). These theories therefore suggest that difficulties with such cognitive processes have downstream effects on the ability to learn to read.

Interestingly, the deficits described in this section have been found across sensory modalities. For instance, the attention span deficit is not limited to the visual domain as an auditory attention deficit has also been observed in adults with dyslexia (see Hari and Renvall, 2001). Additionally, although Sperling et al. (2005) observed reduced perceptual discrimination in the visual domain, similar degraded sensitivity has also been found while tracking sound statistics (Banai and Ahissar, 2018). Most of these theories also seem to relate relatively straightforwardly to statistical learning, with any of these proposed cognitive deficits likely leading to impaired performance on statistical learning tasks. For instance, a deficit of attention, serial order learning, or temporal sampling all could affect statistical learning (at least for those tasks involving serial order or temporal input). Therefore, these theoretical positions provide a rationale not only for poor reading performance but difficulties with statistical learning, explaining any association that might exist between the two.

Another set of theories emphasize limitations in neural function or disturbances to specific neural circuits (Figure 1C), such as the magnocellular system (Stein and Walsh, 1997), the basal ganglia and related circuitry underlying procedural memory (Ullman, 2004; Ullman et al., 2020), and/or the cerebellum (Nicolson et al., 2001). More recently, Nicolson and Fawcett (2019) have suggested that a delay in the development of neural networks associated with communication (e.g., hearing, speech) could eventually cascade into several issues that characterize dyslexia (e.g., phonological circuits that are less organized, a difficulty in automaticity of skills, etc.).

Like the cognitive theories, these neural theories appear to be closely related to statistical learning. Deficits to the magnocellular system and cerebellar and procedural memory systems could all negatively affect statistical learning abilities. On the other hand, each of these theories would provide slightly different predictions for how a statistical learning deficit might play out. Although procedural memory is sometimes thought to be synonymous with statistical learning, the former has a more specific definition focused on the basal ganglia circuitry (Ullman et al., 2020) whereas the latter appears to involve multiple brain circuits such as posterior perceptual networks, anterior brain areas including the prefrontal cortex, in addition to subcortical structures (Frost et al., 2015; Sawi and Rueckl, 2019; Conway, 2020). So, too many tasks considered to measure statistical learning do not necessarily tap into procedural memory; one task that has been implicated as a procedural memory task based on neuroimaging data is the serial-reaction time task (Ullman et al., 2020). Therefore, if deficits on this specific task are found in individuals with dyslexia, then this would be consistent with the procedural memory deficit hypothesis. On the other hand, if deficits are found on other statistical learning tasks that are not thought to tap into procedural memory specifically, such as the embedded patterns task, then this would be indicative of a more general statistical learning deficit. These various tasks will be described in more detail in Section “Statistical Learning.”

All of the previous theories focused mainly on a particular single deficit to account for reading impairment. However, as dyslexia is considered a complex disorder with a multifactorial etiology, it is possible that multi-deficit models will provide improved explanatory power (Elliott and Grigorenko, 2014). Within the multiple deficit model framework, symptoms of complex developmental disorders are thought to be a result of an incremental and interconnected set of dysfunctional processes (Banaschewski and Rohde, 2008; Moll et al., 2020). The multiple deficit theories tend to integrate pre-existing single deficit theories; some examples of this are listed in Figure 1D.

In sum, one common thread among both single and multiple deficit theories is that several of them are focused on difficulties with learning or processing items in a temporal sequence, regardless of whether the input consists of language material (e.g., temporal processing deficit, SOLID, anchoring deficit, procedural learning). Thus, based on a sampling of these deficit theories, it would not be surprising to find that dyslexia is associated with a generalized statistical learning deficit, though as discussed earlier, certain theoretical frameworks provide nuanced views of how such a learning deficit would be manifested. Next, we outline the construct of statistical learning and some common ways that it is measured in the lab.

An inspection of Table 1 reveals that of the 26 separate tasks involving children with dyslexia (across 14 studies), only six showed impairment in statistical learning (He and Tong, 2017; Schiff et al., 2017a; Singh et al., 2018; Tong et al., 2019; Vandermosten et al., 2019; Hedenius et al., 2020), whereas the other 20 reported intact learning (Staels and van den Broeck, 2015, 2017; Schiff et al., 2017a; Inácio et al., 2018; van der Kleij et al., 2018; van Witteloostuijn et al., 2019, 2021a,b; Vandermosten et al., 2019; West et al., 2019). Interestingly, a majority of the studies with children were administered to native speakers of Dutch (Staels and van den Broeck, 2015, 2017; van der Kleij et al., 2018; Vandermosten et al., 2019; van Witteloostuijn et al., 2019, 2021a,b), of which all showed evidence of intact learning, except for performance on one of the tasks by Vandermosten et al. (2019). Besides Dutch and English (Singh et al., 2018; West et al., 2019), tasks were also administered to native speakers in other languages, such as Cantonese (He and Tong, 2017; Tong et al., 2019), Hebrew (Schiff et al., 2017a), Portuguese (Inácio et al., 2018) and Swedish (Hedenius et al., 2020). Continuing with this preliminary evaluation of studies listed in Table 1, certain task features tended to dominate over others. For example, there were more tasks pertaining to the visual and visual-motor domains versus the auditory-visual (Vandermosten et al., 2019) and auditory-motor (van Witteloostuijn et al., 2019, 2021b). This modality issue is discussed further in Section “Discussion.” In addition, most tasks involved temporally presented input, rather than simultaneous spatial arrays. Of the nine tasks examining SRT learning, only two showed impairment (He and Tong, 2017; Hedenius et al., 2020). However, two other studies that also employed overnight consolidation found intact learning on the SRT task (Inácio et al., 2018; van der Kleij et al., 2018). Thus, while the previous meta-analysis examining the SRT task in dyslexia concluded that a learning impairment does exist and was greater for children than in adults (Lum et al., 2013), based on the studies reviewed here, it is difficult to draw that same conclusion, especially considering the possibility of publication bias (Schmalz et al., 2017).

On a first pass of Table 1, there does not appear to be a clear pattern of results that would explain why certain studies found a statistical learning impairment in children and others did not. To examine this issue further, we elaborate first on the studies with impaired performance followed by a similar review of findings centered on intact performance.

Hedenius et al. (2020), for instance, reported impaired learning measured by RTs using an alternating SRT. They found that performance (i.e., the learning effect) for the dyslexia group did not improve with either time and exposure/practice, or consolidation as was the case for those without dyslexia. In contrast, He and Tong (2017) used a modified SRT, but unlike Hedenius et al. (2020), had two control groups (reading and age matched) to compare with the dyslexia group. He and Tong (2017) found that children with dyslexia exhibited a significant learning effect in the condition with a higher number of exposures compared to the one with lower number of exposures (in which no significant learning effect was found). Thus, learning was only impaired with less exposure to the sequence; additionally, those with dyslexia were overall slower than both reading controls. Also, unlike He and Tong (2017), Hedenius et al. (2020) did not account for overnight consolidation. Another point regarding methodology, is that in He and Tong (2017), a reward scheme was introduced after every block, which might have affected participant motivation across groups, differentially.

In one of few studies examining the effects of task complexity on statistical learning, Schiff et al. (2017a) demonstrated that grammar complexity (or topological entropy; see Bollt and Jones, 2000) matters while comparing children with and without dyslexia. They used a high complexity grammar to investigate whether chunk strength and grammar system complexity would influence grammaticality judgment. Not only was performance below chance on grammatical accuracy but also on classification in those with dyslexia (compared to reading and chronological age matched controls). We return to the issue of task complexity, in Section “Discussion.”

Singh et al. (2018) was the only study in Table 1 with children to incorporate neurophysiological measures (i.e., electroencephalography) to index statistical learning. They found that while children with dyslexia had atypical learning as reflected in the event-related potential waveforms, their behavioral performance was comparable to typical controls. This is an important finding as it shows a potential discrepancy in the assessment of learning via RTs (which may reflect a type of implicit motor learning, see Batterink et al., 2015) vs. recordings from this particular event-related brain signal (which may reflect attention-based learning, see Singh et al., 2017). This finding resonates with the view that statistical learning involves both implicit and explicit/controlled aspects of learning. It also highlights that neural measures of learning may provide additional information not captured solely by behavioral measures.

Tong et al. (2019) reported impaired learning on both a visual embedded pattern triplet task (similar to that used by Arciuli and Simpson, 2011) as well as orthographic awareness (not discussed here). In the embedded pattern learning task, the inability to acquire the contingencies between the ordered non-linguistic, visual stimuli, suggests that dyslexia is not just associated with a phonological deficit, but a more general learning impairment.

A final study showing impairment found that when children with dyslexia were presented with native (Dutch) and non-native (Hindi) sounds, they did not utilize embedded statistical cues to learn phoneme differences, unlike the children without dyslexia, who did (Vandermosten et al., 2019). This resulted in less distinct phonemic categories which thus subsequently (negatively) affected their ability to connect between these sounds and their written representation. These findings could potentially be interpreted by the perceptual noise exclusion theory (Sperling et al., 2005) and/or the anchoring deficit theory (Banai and Ahissar, 2018) both of which suggest poor sensitivity to tracking the distribution of sound statistics in individuals with dyslexia.

Turning now to studies showing intact learning, Staels and van den Broeck (2015) reported no group differences between those with and without dyslexia on a variety of Hebb tasks, a finding that persisted, even after attentional functioning was accounted for. Likewise, no group differences were found with both implicit and explicit versions of the SRT and contextual cueing tasks (Staels and van den Broeck, 2017). In the explicit versions of these tasks, participants were informed about the existence of the statistical regularities but in the implicit versions, no such instructions were provided. Although no learning impairments were observed, the children with dyslexia exhibited slower RTs and lower accuracy overall.

In addition to the high complexity grammar mentioned above, Schiff et al. (2017a) also investigated whether knowledge at exposure could be generalized using a low complexity grammar. They found that both groups were able to use chunk knowledge as well as grammatical rules but to different degrees. Thus, although typical learners benefited mainly from chunk strength (high or low), children with dyslexia were primarily influenced by grammar complexity level, chunk strength being secondary. Taken together with their earlier task (Schiff et al., 2017a; high complexity grammar), the results suggest dyslexia may be associated with differences in learning pattern information, which might not manifest as a learning impairment per se unless using a more complex grammar (see also Schiff and Katan, 2014).

Inácio et al. (2018) measured artificial grammar learning over multiple sessions. During the third session (only), a distraction task was administered after which participants performed a grammaticality judgment task, thus obtaining a measure of grammaticality classification during which participants were instructed to respond ‘quickly.’ No differences in learning were observed, despite overall slower RTs in the dyslexia group. Accuracy scores also showed intact learning. Inácio et al.’s (2018) interpretation of comparable learning across groups was based on: reasonable effect sizes and overall performance. However, apart from group level analyses, an individual (d’) grammatical discrimination index based on classification accuracy revealed that only a few high performers in each group showed good discriminability but the majority, irrespective of group, performed at or below chance. Additionally, there was no significant correlation between this d’ index and the accuracy with which participants could later reproduce the previously viewed sequences using a set of cards with geometric figures. The fact that so many individuals did not appear to show a learning effect could be indicative of issues with the task itself.

In a large set of studies, van Witteloostuijn et al. (2019, 2021a, 2021b) showed that learning was intact across tasks, whether they were SRT, visual statistical learning, or auditory non-adjacent learning. The lack of strong group differences despite presenting stimuli in visual-motor and auditory-motor domains across a variety of tasks is quite telling (van Witteloostuijn et al., 2019, 2021a). Similarly, van der Kleij et al. (2018) also reported no group differences after testing children with both SRT and spatial contextual cueing tasks on each of 2 days. Interestingly, resonating with some of the previously reviewed studies, participants with dyslexia had overall longer RTs. Likewise West et al. (2019), also reported no overall significant group differences in learning on the SRT task. However, they found that the dyslexia group showed initial poor learning relative to controls, though they continued to show learning throughout the task.

Thus, although the majority of studies demonstrated a lack of group differences, more subtle differences in performance were observed, such as overall increased RTs for those with dyslexia compared to typical readers. In other words, the lack of a detectable group difference need not mean that statistical learning is comparable across groups. In fact, slower RTs could indicate a temporal processing or visual-spatial attention deficit, even though overall no statistically significant group differences on the primary measures of learning were observed. It is therefore possible that for some tasks, the ways of assessing learning were not sensitive enough to detect group differences and may only become apparent with more complex input patterns (e.g., Schiff et al., 2017a) or the use of neural measures (e.g., Singh et al., 2017). Individual variability within each group is also important to examine more closely (e.g., Inácio et al., 2018), as pointed out by Arciuli and Conway (2018).

To summarize, the studies reviewed here do not provide strong evidence for a statistical learning impairment in children with dyslexia. However, several potential areas to follow up are the role of stimulus complexity (Schiff et al., 2017a) overnight consolidation (Hedenius et al., 2020), and whether there exists a generalized information processing delay as indicated by longer RTs in some studies (van der Kleij et al., 2018; West et al., 2019). Furthermore, it may be important to not only rely on behavior but also to include neural measures (e.g., Singh et al., 2018; see also for instance Menghini et al., 2006). Finally, more work is needed examining individual differences to help understand heterogeneity of abilities within dyslexia (Inácio et al., 2018). Some of these points will be elaborated upon in Section “Discussion” below, after we review the adult studies.

Of the 19 separate tasks involving adults (across 10 studies), 11 of them showed impaired learning in dyslexia (Bogaerts et al., 2015; Gabay et al., 2015; Kahta and Schiff, 2016, 2019; Henderson and Warmington, 2017; Schiff et al., 2017b; Sigurdardottir et al., 2017; Dobó et al., 2021). The impairments were observed for a variety of task types, including linguistic (both auditory and visual) versions of the Hebb task (Bogaerts et al., 2015; Henderson and Warmington, 2017), linguistic and non-linguistic auditory versions of the SRT task (Gabay et al., 2015), non-linguistic auditory versions of the artificial grammar learning task (Kahta and Schiff, 2016; Schiff et al., 2017b), and a non-linguistic visual embedded pattern task (Sigurdardottir et al., 2017). Thus, in contrast to the findings with children, these studies are more consistent with a global, domain-general statistical learning impairment in adults with dyslexia.

However, intact statistical leaning was also found across task types (Staels and van den Broeck, 2015; Kahta and Schiff, 2016; Henderson and Warmington, 2017; Samara and Caravolas, 2017; Schiff et al., 2017b). Interestingly, intact learning was reported for a majority of studies involving a simultaneous presentation in a finite state grammar paradigm (Kahta and Schiff, 2016; Samara and Caravolas, 2017). Studies with adults were administered in many different native languages, though no one language was predominantly represented as was the case with children (e.g., Dutch). For instance, adult participants spoke either Dutch (Bogaerts et al., 2015; Staels and van den Broeck, 2015), English (Gabay et al., 2015; Henderson and Warmington, 2017), Hebrew (Kahta and Schiff, 2016, 2019; Schiff et al., 2017b) or Icelandic (Sigurdardottir et al., 2017).

As in the previous section, we elaborate first on the studies with impaired performance followed by a similar review of findings showing intact performance. Henderson and Warmington (2017) found partially impaired (significant initial effect that attenuated over time) statistical learning via an auditory-linguistic Hebb repetition task. Additionally, to rule out whether any observable group differences were due to poor motor control, Henderson and Warmington (2017) also administered a motor tapping (non-sequential) task to all participants and concluded that those with dyslexia did in fact show a significantly slower processing speed than adults without dyslexia.

In another set of Hebb learning studies, Bogaerts et al. (2015) reported that adults with dyslexia initially exhibited impaired learning and need significantly more exposure to repetitions (than controls) on a Hebb task to reach a pre-established learning criterion. Some of these adults, however, (although not all) eventually developed long-term serial order representations, over time. Adults with dyslexia also had poorer (relative to controls) baseline performance on fillers, suggesting worse memory capacity, for serial order. Retention (in terms of memory savings over time), however, became comparable across groups after 1 month of initial learning. Even after controlling for baseline differences across groups (average filler performance), the finding of impaired serial order learning persisted. Additionally, once learned, though initially slow, both groups showed comparable retention on Hebb sequences, after consolidation (of 1 day; and even after 1 month). In their second experiment (Hebb learning and pause detection), Bogaerts et al. (2015) sought to replicate their first experiment but they also measured lexical engagement by introducing lexical competition in Hebb sequences via a pause detection task (for a more nuanced description see Szmalec et al., 2012 and Bogaerts et al., 2015). To measure lexical engagement, lexical competition is expected to occur when participants are presented with Hebb sequence containing base words (e.g., lavabu) that would interfere with words form the participants’ (Dutch) mental lexicon (e.g., lavabo, meaning kitchen sink). Thus, participants when presented with these sequences, both with and without an artificially embedded pause had to quickly indicate the presence (or not) of a pause. Successful detection of the pause would mean selecting from amongst multiple candidates in the mental lexicon, depending on knowledge of recently learned Hebb sequences (containing lexical neighbors). Bogaerts et al. (2015) concluded that lexical engagement (interaction of new items with pre-existing word-forms in the mental lexicon) did occur for control participants but was less robust for adults with dyslexia, even after 1 month, although this effect was not strongly (statistically) supported. When combined, results indicate that adults with dyslexia were impaired in the long-term learning of verbal serial order information and also showed weaker lexicalization of novel word forms.

Notably, the predecessor to both Szmalec et al. (2011), Bogaerts et al. (2015), and Henderson and Warmington (2017), who also showed Hebb learning impairment for individuals with dyslexia, administered three Hebb versions: a verbal-visual, verbal- auditory and visuospatial version; and reported poor learning amongst those with dyslexia for all three task versions. These findings in adults contrast with the relatively intact Hebb learning in children. It is thus possible that there may be developmental and cognitive changes in adulthood that account for the impaired Hebb learning across sensory domains.

In another study indicating impaired learning, Gabay et al. (2015) incorporated speech syllables used by Saffran et al. (1996) that were presented to participants during a passive familiarization. At test, they then had to respond by judging which one of a pair (word and part-word) was similar to sounds in familiarization, via a two alternative forced choice response option. Transition probabilities varied from 1.0 within a word (hearing da after bi in bidaku) to <1.0 between words (probability of encountering any of the beginning syllables in padoti, gulabu or tupiro presented after ku in bidaku goes down to 0.33). Similarly, they also had a non-speech version of the task in which they used non-linguistic tones that were from Saffran et al. (1999). After continuous temporal presentation (exposure) of the two types of stimuli, participants then had to complete a familiarity judgment at test. Relative to the control group, Gabay et al. (2015) reported impairment for those with dyslexia, across both speech and non-speech materials, despite overall above chance performance.

Similarly, Sigurdardottir et al. (2017), used an embedded pattern task with a spatio-temporal presentation, but resorted to offline, rather than online-motor response acquisition. Sigurdardottir et al.’s (2017) findings were nonetheless similar to those of Gabay et al. (2015), revealing that unlike typical learners, those with dyslexia had trouble with familiarity judgment (at test) of shape pairs (nonsense letter-like stimuli) they had viewed previously during familiarization. Thus, the nature of the learning impairment appears consistent across linguistic and non-linguistic patterns, at least for tasks with embedded patterns.

Turning to a visual artificial grammar paradigm, Kahta and Schiff (2016) also reported group differences. Performance of those without dyslexia exceeded chance under both implicit and explicit learning conditions (where participants were or were not informed of embedded rules at training). However, in the dyslexia group, impairment was reported for implicit learning but was intact on the explicit version of the task. Keeping the same paradigm but changing the sensory modality, Schiff et al. (2017b) investigated the effects of feedback on implicit and explicit learning with auditory stimuli. Despite a decrease in false alarms across groups, feedback under implicit conditions lead to increased hits only for those without dyslexia; however, under explicit conditions, feedback was more helpful to those with dyslexia.

Kahta and Schiff (2019) similarly investigated artificial grammar learning under implicit and explicit conditions, using temporally presented tones. They found that results in the auditory modality were similar to the implicit condition visual findings of Kahta and Schiff (2016), in that although both groups performed above chance, d’ values were significantly lower for those with dyslexia compared to controls, linking such performance to an anchoring deficit (Ahissar, 2007). On the other hand, auditory modality findings showed comparable performance across groups under explicit learning conditions. Together, these findings suggest a complex interplay between the presence of instructions and feedback and how these factors interact to influence the learning of visual and auditory patterns dyslexia. More work is needed to clarify these research findings.

In addition to the studies above, only one study examined statistical learning in adolescents (Dobó et al., 2021), with a report of impaired learning for the dyslexia group on an auditory statistical learning task. More work is needed to follow-up on a potential auditory statistical learning impairment in adolescents. Currently, it is difficult to conclude based on this study alone, that adolescents with dyslexia are impaired on statistical learning.

Turning next to studies showing intact learning, two of these are already discussed above in tandem with the impaired findings. In short, Kahta and Schiff (2016), found impaired learning in the implicit version of their artificial grammar learning study, compared to relatively intact explicit learning. Similarly, Schiff et al. (2017b), in addition to the points above, showed that children with dyslexia had more false alarms than controls under implicit learning but they did benefit more from feedback in the explicit condition.

Samara and Caravolas (2017) concluded that adults with dyslexia learned grammaticality and chunk strength knowledge that was comparable to typical readers on both a linguistic as well as a non-linguistic version of the same artificial grammar learning task. Notably, Samara and Caravolas (2017) explicitly instructed participants to both memorize, as well as recall each letter string following exposure, which may have impacted learning outcomes. Additionally, because Samara and Caravolas (2017) excluded some participants based on their clinical profile (due to comorbid diagnoses), their study was unable to ascertain if a more heterogeneous group of individuals would have been more likely to show impaired learning. One thing to note is that the artificial grammar learning paradigm typically only incorporates offline measures of learning, which may mask more subtle differences in the learning profiles between typical and atypical readers that could only be detected with online measures.

In addition to the impaired Hebb and manual-motor learning described above, Henderson and Warmington (2017) also found overall intact learning in a SRT. Further, they reported that the dyslexia group tended to have longer RTs and more variability compared to controls, which echoes other indications of increased RTs in children, as reviewed in the previous section. Overall, both groups showed evidence of a learning effect in terms of RTs and accuracy (significantly higher for sequenced vs. random trials) as well as in terms of consolidation (both groups showed sustained learning that grew stronger with time).

In line with their findings in children, but in contrast with previous work on Hebb learning impairments in dyslexia (e.g., Szmalec et al., 2011; Henderson and Warmington, 2017). Staels and van den Broeck (2015) did not report Hebb learning deficits for adults with dyslexia in any modality (verbal-visual, verbal-auditory, non-verbal visuo-spatial).

To summarize, compared to research on children, relatively more tasks found learning impairments in adults with dyslexia, across a variety of task types and learning conditions. This result is in contrast to two previously described meta-reviews which concluded there was a greater impairment in children with dyslexia relative to adults (Lum et al., 2013; van Witteloostuijn et al., 2017). However, whether the current findings of learning impairment in adults is an actual effect or due to publication bias remains to be seen, and is an issue that will be taken up in greater detail next.