Literacy is an exquisite example of human ingenuity. Written scripts are composed by graphs¹, i.e., artificial two-dimensional geometric-like shapes (cf. Chang et al., 2018) that are arbitrary but, when learning to read, become visual counterparts of linguistic units as phonemes, syllables, or morphemes (e.g., letters in the Latin alphabet, kanas and kanjis in Japanese), and gears of written words (e.g., Pelli et al., 2003; Grainger, 2018). Reading thus bridges visual object recognition and language.

Along with reading development, a hierarchically organized, orthographically tuned circuitry is built along the visual ventral stream, originally dedicated to visual recognition of familiar objects (for a review, see Dehaene et al., 2015). Skillful graph and visual word recognition depend on fast access to abstract orthographic representations (usually called abstract or symbolic letter identities) which are not determined by physical (visual, low-level), phonological, or motor similarity and are underpinned by the left ventral occipitotemporal cortex, vOT (Dehaene et al., 2005; Rothlein and Rapp, 2014). Readers’ immunity to physical differences of allographs (visual forms of the same graph) is found in different scripts: e.g., in Latin alphabet, A and a; in Japanese kanas, and ; in Arabic abjad, and (e.g., Bowers, 1996; Carreiras et al., 2013; Kinoshita et al., 2019). It depends on long-term changes in the perceptual space of graphs, which are consequence of learning to read. For example, B is physically equidistant to both p and b, and hence, without any experience, observers would be as fast in discriminating the pair B-p as the pair B-b. However, readers are slower in discriminating the pair B-b in consonance with the degree of perceptual similarity of graph representations (Lupyan et al., 2010; for recent evidence with training on an artificial script, see, Wiley and Rapp, 2021). This example demonstrates that learning to read is an ecological example of perceptual learning (Gibson, 1969, 1970; Dehaene et al., 2005; Goldstone and Byrge, 2015).

Probably because learning to read puts heavy demands on visual processing, the observation that motor training via handwriting, HW (that is, writing by hand) benefits more subsequent visual graph recognition than other learning experiences has been startling since the earliest studies (Jeffrey, 1958; Williams, 1969, 1975; Jensen and King, 1970; Koenigsberg, 1973). The advantage from HW training when learning visual graphs is robust. It is found in different written systems and types of scripts (for a recent meta-analysis, see Araújo et al., 2021; e.g., Jensen and King, 1970; Guan et al., 2011; Cao et al., 2013b; Li and James, 2016; Xu et al., 2020; Vinci-Booher et al., 2021). It is especially strong for highly confusable graphs which share visual features and in the most extreme the whole shape, differing only by orientation, such as mirror images (e.g., d and b) or rotations in the image plane (i.e., plane rotation: e.g., N and Z, or n and u) (Hendrickson and Muehl, 1962; Williams, 1969, 1975; Torres et al., 2020). It is found relative to visual-only (e.g., looking at) and to motor control (e.g., typewriting on a keyboard, pointing to, circling) training, regardless of training in phonological correspondences, adoption of natural or artificial scripts, amount of training (single vs. multiple training sessions), and age of learners (Araújo et al., 2021; for an overview, see James, 2017; e.g., Williams, 1969; Longcamp et al., 2006, 2008; James, 2010; Bara and Gentaz, 2011; Guan et al., 2011; Suggate et al., 2016; Labat et al., 2020; Mayer et al., 2020; Seyll et al., 2020). Note that control training often leads to improvement in subsequent visual graph recognition as well, but HW training usually excels it (Longcamp et al., 2006; Kiefer et al., 2015; Labat et al., 2020; Mayer et al., 2020; Fernández-López et al., 2021). This benefit from HW is impressive because it corresponds to a transfer effect (Gilbert and Li, 2012): performance in an untrained (non-motor) visual recognition task on the new graphs is enhanced via HW training, indicating that neural plasticity is not restricted to the brain underpinnings of the graphomotor task but extends to those of visual graph recognition at the left fusiform gyrus in the vOT (Longcamp et al., 2008; James, 2010).

It is consensual that writing by hand, stroke by stroke, establishes a connection between the visual percept of the graph and the motor plan for creating it, resulting in a sensorimotor experience that influences learning to read. Most research has provided descriptive insight only, occasionally legitimating evidence-based programs (e.g., Bara et al., 2016; Mayer et al., 2020; Torres et al., 2020) given that HW is a worldwide strategy in literacy instruction (e.g., Tan et al., 2005; Bara and Gentaz, 2011; Itaguchi et al., 2015; Xu et al., 2020). However, this benefit from HW training does not necessarily imply that knowledge on how to write the graphs (that is, on motoric representations) is the underlying mechanism. This (premature) conclusion confuses the expression of a learning experience with the underlying cognitive representations and processes (Norris and Cutler, 2021).

In fact, explanatory insights have been rare (notwithstanding recent exceptions: Li and James, 2016; Zhai and Fischer-Baum, 2019; Seyll et al., 2020; Vinci-Booher et al., 2021; Wiley and Rapp, 2021). We thus came to have misgivings about the underlying mechanism: Why does HW training benefit visual graph recognition? Is it due to perceptual learning and visual segmental analysis regardless of the motor act (e.g., Williams, 1975; Courrieu and de Falco, 1989; Seyll et al., 2020)? Or does it depend on internal motor simulation (e.g., Longcamp et al., 2008, 2016) such as that of stroke order during visual graph recognition (e.g., Parkinson and Khurana, 2007; Itaguchi et al., 2015)?

It is time to put the phenomenon into perspective, confronting potential explanations, in order to bring to light the underlying cognitive mechanism. In the present work, we critically review evidence from pioneering research (e.g., Hendrickson and Muehl, 1962; Pick, 1965; Williams, 1969; Jensen and King, 1970; Koenigsberg, 1973) to more recent functional resonance imaging (fMRI) studies (e.g., Longcamp et al., 2006; James and Atwood, 2009; James, 2010; Cao et al., 2013b; Vinci-Booher et al., 2021). Our aim is to aggregate this research, adopting a theoretically based perspective that has hitherto been largely missing. Specifically, we consider the two alternative theoretical frameworks (about conceptual representations, and hence, not exclusive of letter representations) that can be associated with most research on the benefit from HW training even if this theoretical discussion has often been dismissed (e.g., Bara and Gentaz, 2011; Labat et al., 2020; Torres et al., 2020). The debate between these two theoretical perspectives is about the nature of cognitive representations, that is, about their format (i.e., the nature of the code used to represent information) and not about their content (that is, which information is stored; for a critical discussion, see, e.g., Glenberg et al., 2013; Barsalou, 2016; Goldinger et al., 2016; Machery, 2016; Mahon and Hickok, 2016).

On the one hand, embodied cognitive accounts (also called grounded or situated cognition: e.g., Allport, 1985; Barsalou, 2008; Glenberg et al., 2013) hold that the content and the format of cognitive representations is isomorphic: sensory concepts have a sensorial format and action concepts have a motor format. Therefore, representations of graphs and of written words would not be symbolic but rather sensorimotor. During visual graph recognition, the graphic motor programs acquired via HW would be reactivated or simulated because they would be an inner part of graph representations. This sensorimotor mechanism would be responsible for the benefit from HW training in visual graph recognition (e.g., Bara and Gentaz, 2011; Longcamp et al., 2016; Labat et al., 2020; Xu et al., 2020). Action-perception coupling via HW training would be critical for the development of the reading-specialized circuitry (Longcamp et al., 2008; James and Atwood, 2009) because reading would involve a gesture decoding system, located within a region of the left dorsal premotor cortex, PMd (Brodmann Area 6; e.g., Longcamp et al., 2008; Nakamura et al., 2012; Cao et al., 2013b).

On the other hand, abstractionist, symbolic accounts (e.g., Goldinger et al., 2016; Machery, 2016; Mahon and Hickok, 2016) do not deny the role of sensorimotor experiences when learning to read but hold that abstract representations are not reducible to, and hence differ from, modality-dependent ones. In what regards visual graph processing, ample behavioral, neuroimaging, and neuropsychological evidence (e.g., Rapp and Caramazza, 1997; Dehaene et al., 2005; Lupyan et al., 2010; Dufor and Rapp, 2013; Wiley and Rapp, 2021) shows that symbolic, amodal graph identities are core representations in reading and writing in alphabetic and non-alphabetic scripts (e.g., Carreiras et al., 2013; Rothlein and Rapp, 2014; Kinoshita et al., 2019). These abstract orthographic representations are connected via bidirectional links to the input (visual) and output (motor) systems, with automatic spreading of activation between them (Rapp and Caramazza, 1997; Dufor and Rapp, 2013). HW training would thus assist on the emergence of abstract orthographic representations due to activation dynamics (that is, because activation cascades automatically) and not because motoric representations were an inner part of orthographic representations. In this vein, HW training would benefit subsequent visual graph recognition due to the operation of a perceptual learning mechanism, resulting in long-lasting changes in the perception of the trained graphs (e.g., Gibson, 1969, 1970; Williams, 1975; Courrieu and de Falco, 1989; Goldstone and Byrge, 2015; Seyll et al., 2020).

Both frameworks predict that HW training benefits more subsequent visual graph recognition than other learning experiences. However, they disagree about the nature of the representations and about the putative mechanisms underpinning the HW benefit. In this work, we highlight some of the loose ends that research has left and present our theoretical framework and hypotheses. Before our proposal, we start by pinpointing the cognitive components involved in HW, to then discuss three promising theoretical accounts framed by the symbolic cognitive framework (i.e., the perceptual variability and the visual analysis hypotheses) and by the embodied cognitive framework (i.e., the stroke processing hypothesis). For each hypothesis, we first present critical positive evidence and next the evidence that questions it. Note, however, that it is not our aim to provide an exhaustive literature review but one that is unbiased and tackles the critical evidence for the present discussion. We also detail the predictions that follow, because, like the devil, unveiling the underlying cognitive mechanisms is on the details.

Learning to read is often accompanied by proxies of HW as copying and tracing (Kiefer et al., 2015; Mayer et al., 2020). It is a multisensory experience bridging visual graphs with speech sounds and motor gestures (Pegado et al., 2014). Consequently, letters have multiple codes (i.e., visual, motor, phonological) and are involved both in reading and in writing, which in turn also comprise multiple types of representations (Abbott and Berninger, 1993; Rothlein and Rapp, 2014). Therefore, the cognitive mechanisms and representations involved in a transfer effect from HW training to visual graph recognition are not straightforward. It is undeniable that HW training has effects in writing and might also have in other abilities beyond the written domain (Abbott and Berninger, 1993). However, the scope of the present work is on the benefit from HW training on subsequent visual recognition of graphs. We are isolating a specific challenge that is posed when learning to read, that is, the emergence of abstract graph identities, which can be distinguished from other challenges posed in orthographic processing during reading development (e.g., transposed letter effect; phonological consistency).

The first strategy to enlighten the transfer effect from HW training to visual graph recognition is by process decomposition. The rationale here is that of the perceptual expertise literature (Curby and Gauthier, 2010): the training task is a vehicle for encouraging differences in processing and/or in representation of graphs, and transfer effects from training to testing tasks depend on the common denominators (Gilbert and Li, 2012). Such rationale is not only optimal to theory testing but can also readily translate into education. By identifying the key components of this benefit, other tasks besides HW can be used to optimize learning to read. Furthermore, given that HW is becoming an obsolete task in the digital era, its inclusion in school activities depends on the contribution of HW to other facets of literacy beyond writing per se (Wiley and Rapp, 2021). For unveiling this cognitive mechanism, we need to isolate the common denominators, that is, which representations and processes in the learning experience are critical to subsequent visual graph recognition and reading. To this aim, we first consider the cognitive components involved in HW.

Handwriting is a perceptual-motor multi-component task that involves a plethora of processes, including balance, eye-hand coordination, focused attention, visual processing, fine coordination of hand movements, and precise motor control of spatial and temporal constraints (Waterman et al., 2015; Julius et al., 2016). Behaviorally, HW is characterized by legibility and fluency (i.e., accuracy and speed in reproduction). It is underpinned by a frontoparietal associative striatum-cerebellar circuitry that also engages attention and executive processes (Makino et al., 2016; Palmis et al., 2017; Vinci-Booher et al., 2019). Therefore, at first sight, multiple possible mechanisms could be involved in a transfer effect to visual graph recognition and reading (for a similar discussion, see, Wiley and Rapp, 2021).

Notably, the temporal course of HW learning reveals two aspects that could be key for unveiling the underlying cognitive mechanism. Like other forms of motor learning, HW follows a well-characterized temporal course of two phases, both dependent on the cortical-cerebellar loop (the primary motor, somatosensory, dorsal premotor, and parietal cortices, the thalamus and cerebellum) and the cortico-striatal loop (composed by the same motor cortical areas, the thalamus and basal ganglia; Makino et al., 2016; Palmis et al., 2017). It begins with a fast-learning phase of rapid improvement but with slow and highly variable graph reproduction. Movements are overly guided by the visual stimuli, with exploration of multiple motor behaviors, contingently to the resulting visual output (Maldarelli et al., 2015). Next, in a slow-learning phase, refinement of graph reproduction occurs over a longer time course. HW becomes legible, fluent, and highly stereotyped but still relies more on visual feedback than proprioceptive (tactile and kinesthetic) feedback (Palmis et al., 2017). Thus, the two key aspects are that, during the first phase, HW is accompanied by highly variable visual outputs, and hence, perceptual variability of the graphs to-be-learned is maximum, and in both learning phases visual processing is prevalent.

This temporal course also highlights the first loose end of this topic of research. Most studies on the benefit from HW training when learning visual graphs likely tap mostly into the first phase of motor learning. These studies were usually short-term, many with a single training session of 1–3 min (e.g., Naka and Naoi, 1995; Suggate et al., 2016) to 20–25 min (e.g., Jensen and King, 1970; Guan and Wang, 2017) and subsequent visual graph recognition was tested immediately or 24 h post-training (e.g., Williams, 1969; James and Atwood, 2009). From the 50 studies meta-analyzed by Araújo et al. (2021), half had just one or two training sessions (of less than 30 min) and only six studies had more than seven training sessions (e.g., Kiefer et al., 2015; Mayer et al., 2020). Follow-up testing has been rare, although post-training gains are kept after at least one week (Longcamp et al., 2006, 2008; Cao et al., 2013b; Seyll et al., 2020; Vinci-Booher et al., 2021). Note, however, that when children were trained on real letters, uncontrolled post-training exposure could have been involved in the follow-up results (e.g., Longcamp et al., 2005). More important to the present work, in these training studies HW is always preceded, accompanied, and followed by visual graph processing, and hence, visual perceptual processes seem to be critical. This observation agrees with fMRI evidence showing enhancement of functional activity in occipitotemporal regions for HW with ink vs. without ink (Vinci-Booher et al., 2019). Therefore, in short-term training studies, HW training is contingent to visual pattern learning, and hence, its benefit could be due to perceptual learning of graphs (Williams, 1975; Courrieu and de Falco, 1989).

The first two hypotheses discussed here, that is, the perceptual variability (James and Engelhardt, 2012; Li and James, 2016) and the visual analysis (Koenigsberg, 1973; Courrieu and de Falco, 1989; Seyll et al., 2020) hypotheses are framed by symbolic accounts of cognitive representations (e.g., Goldinger et al., 2016; Machery, 2016; Mahon and Hickok, 2016). Both were originally proposed in studies with preliterate children, and hence, with learners that had no previous reading expertise in any script (we return to this point in Section “The nature of graph representations”). These hypotheses are not mutually exclusive but focus on different operations of perceptual learning. As aforementioned, this mechanism is responsible for the emergence of symbolic graph identities (Gibson, 1969; Dehaene et al., 2005; Goldinger et al., 2016) as expressed by faster discrimination of the pair B-p than B-b by Latin-alphabet readers (Lupyan et al., 2010). In fact, this example illustrates the two perceptual challenges that a learner faces, each one emphasized by one of the two symbolic hypotheses discussed below.

Stroke representations are not necessarily motor, gestural in format. In fact, sensorimotor representations are already a transduction (Machery, 2016). The emphasis by Rapp and colleagues (e.g., Rapp and Caramazza, 1997; Dufor and Rapp, 2013; Rothlein and Rapp, 2014) in the abstract, amodal quality of these representations highlights their symbolic nature, which gives them computational fitness (Mahon and Hickok, 2016). Indeed, the PMd, sometimes referred as a motor center (Roux et al., 2009; Longcamp et al., 2016), is involved in serial sublexical orthographic processes shared by writing and by pseudoword reading (Pattamadilok et al., 2016). From the reviewed evidence, it is clear that the benefit from HW is not about the motor act itself (e.g., Courrieu and de Falco, 1989; Xu et al., 2013, 2020). Our proposal is that the benefit from HW training when learning visual graphs is about computations over symbolic representations (for a similar proposal and supporting empirical evidence, see Wiley and Rapp, 2021). Whether these regard strokes (units of movement, primitives of HW), visual features (image components, primitives of visual object recognition), or both must first be seriously discussed. Several authors have pointed out that visual analysis and dynamic movement could be involved (e.g., Courrieu and de Falco, 1989; Cao et al., 2013a; Merritt et al., 2020; Vinci-Booher and James, 2020), but few studies tried to disentangle them.

Cognitive science has recently shown a renewed interest in the role of HW training when learning to read (for an overview, see, James, 2017). Indeed, a large body of empirical evidence supports the advantage from HW training relative to control training in subsequent visual graph recognition (Araújo et al., 2021; for an overview, see, James, 2017; e.g., Williams, 1969; Longcamp et al., 2006, 2008; James, 2010; Bara and Gentaz, 2011; Guan et al., 2011; Suggate et al., 2016; Labat et al., 2020; Mayer et al., 2020; Seyll et al., 2020). However, the nature of the underlying cognitive mechanism has been elusive and rarely addressed (Gibson, 1970; Williams, 1975; Kaufman, 1980; Li and James, 2016; Zhai and Fischer-Baum, 2019; Mayer et al., 2020; Seyll et al., 2020; Vinci-Booher et al., 2021). The available theoretical proposals have hitherto been unspecified. Therefore, it was not fully clear which predictions would follow and which patterns of performance would empirically distinguish them. In this work, we presented the most promising theoretical accounts, detailed their predictions, and critically revisited key empirical evidence.

We join other authors (e.g., Gibson, 1970; Kaufman, 1980; Courrieu and de Falco, 1989; Mayer et al., 2020; Seyll et al., 2020) in the proposal that HW training is a vehicle for perceptual learning of visual graphs. Visual segmental analysis would be the key element in HW training by highlighting diagnostic features of visual graphs which then would assist the emergence of perceptual representations to be involved in subsequent visual graph recognition (see also Wiley and Rapp, 2021).

Although beyond the scope of the present work, the theoretical proposals discussed here have implications for the nature of mental representations. Embodied and symbolic cognitive accounts are two perspectives with dramatically different approaches in this regard. Note that both proposals are able to accommodate the available fMRI evidence (e.g., Longcamp et al., 2003, 2008; James and Atwood, 2009; James, 2010; James and Engelhardt, 2012; Nakamura et al., 2012; Vinci-Booher et al., 2016, 2021; Vinci-Booher and James, 2020) but they do differ on the cognitive mechanisms responsible for such effects. According to the embodied sensorimotor accounts, the fronto/parietal regions within the writing network are activated when viewing graphs or written words because these regions underpin motoric representations that are core of orthographic (sensorimotor) representations (e.g., Parkinson and Khurana, 2007; Longcamp et al., 2008, 2016; James and Atwood, 2009; Cao et al., 2013b; Itaguchi et al., 2015). According to symbolic accounts, the observed motor activation is rather due to information spreading throughout the orthographic system; it is coactivation, not causation (e.g., Rapp and Caramazza, 1997; Rothlein and Rapp, 2014; Mahon and Hickok, 2016). In fact, from the reviewed literature, we must conclude that there is no compelling evidence that embodied representations are necessary for understanding the benefit from HW training in visual graph recognition (for a meta-analysis and discussion, see, Araújo et al., 2021). In fact, the recent results of Wiley and Rapp (2021) show that HW training leads not only to the emergence of motoric representations but also of dissociable symbolic orthographic representations.

The present work has also raised several questions to be addressed in future research. We thus present further predictions and future directions that could be enlightening in what regards the mechanism underpinning the benefit from HW training when learning visual graphs. First, although the idea of a perceptual learning mechanism via visual analysis seems parsimonious, it is mainly corroborated by behavioral evidence showing that this type of visual training leads to either similar or even larger post-training gains than HW training (e.g., Williams, 1969, 1975; Koenigsberg, 1973; Courrieu and de Falco, 1989; Li and James, 2016; Seyll et al., 2020). In the present work, we discussed how this visual analysis hypothesis could explain prior evidence which did not test it nor considered it. We also presented the arguments in favor of this proposal, considering detailed predictions. However, we must acknowledge that even if the behavioral effects are the same, the benefit from visual segmental analysis and HW trainings might result from different neurocognitive mechanisms. Cognitive, neural, and behavioral changes when learning visual graphs will be closed intertwined and all are relevant for understanding the underlying mechanism. At the cognitive level, we believe that the predictions presented here, especially those regarding the opposition between stroke and visual feature processing, will be especially revealing (see Table 1). In what regards the neural implementation, a promising technique to address the underlying mechanism is multivariate pattern analysis of brain responses, especially when adopted with high-temporal resolution techniques as EEG (King and Dehaene, 2014). Such classification algorithms could assist in accomplishing three aims: (i) determining the temporal course of emerging representations of graphs in the brain; (ii) testing whether the brain pattern of response is able to predict post-training gains in visual graph recognition; (iii) testing whether one can predict which type of training the participants were in (e.g., HW vs. visual analysis) based on patterns of brain responses to visual graphs after training. Achievement of these aims would be especially revealing on whether the similar benefits from HW training and visual segmental training (e.g., Courrieu and de Falco, 1989; Seyll et al., 2020) do reflect the operation of the same neurocognitive mechanism and on whether top–down effects related with a late stage of processing would be involved in the benefit from HW training (see Table 1). In this regard, convergence across methods will provide a better characterization of the components involved. This line of research is thus relevant to the future development of our framework.

Second, we also discussed two loose ends of this topic that hopefully will be considered in future research. We are hardly the first to consider them, for which the earliest studies contributed with deep insights (e.g., Tawney, 1972; Williams, 1975; Kaufman, 1980; Kirk, 1980), albeit these pioneer works have (surprisingly) been underestimated by recent research. On the one hand, training regime (i.e., total amount, frequency, and duration of training) has been overlooked (although with references by Jeffrey, 1958; Longcamp et al., 2006), probably because the benefit from HW has been found both in single and multiple session training (e.g., Williams, 1969; Li and James, 2016; Mayer et al., 2020; Seyll and Content, 2020). However, the nature of graph representations may change as a function of training regime, which would agree with the different patterns of fMRI activation found when children (6-year-old beginning readers and 8-year old) and adults (fluent readers) viewed letters (Vinci-Booher and James, 2020) and the observation that functional connectivity between visual and motor brain regions found immediately after HW training were already gone after one week (Vinci-Booher et al., 2021).

Notably, a manipulation of training regime could also be key in hypothesis testing. If the benefit from HW training were about stroke processing, it would take time to develop, and hence, the benefit relative to control training would increase along sessions. Alternatively, if the benefit from HW training is due to visual analysis, then the largest difference relative to control training would occur early on, which would dissipate with stabilization of graph representations. Nonetheless, whenever testing of visual graph recognition occurs multiple times along training, then the contribution of testing for the learning curve must be controlled to ensure that it is not confounded with the independent contribution of training (for this kind of strategy, see, Wiley and Rapp, 2021). Training along multiple sessions also allows for the involvement of sleep, which is relevant in perceptual learning and visual discrimination (Stickgold et al., 2000). Indeed, sleep has a significant role in the stability of post-training gains in mirror-image discrimination of graphs after HW training (Torres et al., 2020). In this vein, follow-up assessment is of critical importance (Longcamp et al., 2005, 2008; Vinci-Booher and James, 2020; Vinci-Booher et al., 2021). However, very few studies included a follow-up and with a disperse interval, from one week to several months. When the new graphs were real letters, these studies also had the possible confound of uncontrolled post-training exposure (e.g., Longcamp et al., 2005).

On the other hand, when learners are already experts in reading and HW in their first script, a literacy-specific network is already established. Therefore, they might rely on it when learning new graphs. The research has suggested, however, that this might not be the case. Indeed, neurocognitive and behavioral effects of leaning a new script are similar in (literate) adults and (preliterate) children, even though besides reading skills there is also an age/maturation confound in this comparison (James and Atwood, 2009; Brem et al., 2010; James, 2010; James and Engelhardt, 2012; Brem et al., 2018; Vinci-Booher et al., 2021). Additionally, the earliest studies have examined possible generalization effects from the graphs trained via HW to novel, untrained graphs, and found that such generalization occurred and assisted subsequent learning of letter-sound correspondences (Pick, 1965; Tawney, 1972; Nelson and Wein, 1974). However, since then this generalization effect has not been considered, leading to several questions. For example, could generalization effects depend on the type of diagnostic features of the new script and whether these features resemble or not those of the first script? And if this were the case, with which time course?

These questions also highlight the link between the studies on the benefit from HW training when learning visual graphs and the field of perceptual expertise. Indeed, perceptual expertise can be contrasted to other forms of perceptual learning as it is characterized by robustness and generalization to new contexts and to new items within the expertise domain (Curby and Gauthier, 2010). This field has systematically focused on the interaction between participant (naïve vs. expert), item (novel vs. old), and task demands (individuation vs. categorization), which are all known to be important in development and in expression of expertise.

In fact, the benefit from HW training when learning visual graphs is, in our view, a paradigmatic example of perceptual and of motor expertise. Note that, in literate adults, HW is highly automatic; it consists of rapid sequences of short movements with fast changes in direction, resulting in high-quality, stable, and consistent reproduction of graphs. It is clearly an ecological example of motor expertise like playing a musical instrument (Palmis et al., 2017; Calmels, 2020). Therefore, the convergence with research from perceptual and motor expertise could be fruitful for a deeper understanding of the cognitive mechanism underpinning the benefit from HW training in visual graph recognition (see, e.g., Folstein and Monfared, 2019). This research is also relevant for the expertise literature, given that it is easier to study expertise in a domain with many participants and where relevant stimuli is widely available.

In summary, the investigation of the cognitive mechanism underpinning the benefit from HW training when learning visual graphs goes beyond its realm. It can provide insights into the principles, limits, and possibilities of learning of cultural activities. It is also relevant for generation of testable hypotheses about interactions between training modes and performance benefits. In a broader scope, it also speaks to the nature of cognitive representations. Finally, it has relevance for Education and public policy because it can translate in better designing of literacy programs. When learning visual graphs, HW involves multiple components. Thus, we can no longer advocate the adoption of HW as a holistic school activity. The critical processes can and should be separated in order to implement the best educational strategies in literacy instruction.

Publicly available datasets were analyzed in this study. This data can be found here: OSF Preprints. doi: 10.31219/osf.io/quywt; https://osf.io/wzdxn/.

TF and SA conceptualized together this work, critically revised the manuscript, and prepared the revised version. TF has the first authorship and wrote the first draft of the manuscript. Both authors approved the final version submitted and the revised version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.