From neural matter to rapid symbolic learning in brains and artificial neural networks: a brief overview and perspective

1 Introduction

One of the main features that sets humans apart is the surprisingly rapid acquisition of a vast vocabulary of meaningful symbols in early ontogeny. By around 12 months, most infants begin to produce their first word, and towards 18 months, when vocabulary reaches a certain threshold, growth often accelerates, with children acquiring on average up to 10 words per day, while their receptive vocabulary is likely even larger (e.g., Bloom 2002; Reznick and Goldfield 1992). By adulthood, vocabulary size can reach approximately 42,000 words or more (Brysbaert et al. 2016; see also Anglin et al. 1993).^[1] How can such spectacular growth of a lexicon be explained?

While a range of linguistic, cognitive, and socio-pragmatic factors are known to contribute to early word learning (Goldstein et al. 2003; Samuelson and McMurray 2017; Tomasello 2003), one mechanism that has received considerable empirical attention is fast mapping. This process refers to the ability to form associations between a novel word and its referent after only minimal exposures or sometimes even after a single encounter. Fast mapping was first formally described by Carey and Bartlett (1978), who showed that children, when presented with a blue tray and an olive tray and instructed “Bring me the chromium tray, not the blue one, the chromium one”, successfully inferred that the novel label chromium referred to the olive-colored tray (for a review, see Carey 2010). Subsequent research has examined the robustness of rapid mappings over time in both children and adults, showing successful retention of a single new word even after one month (Markson and Bloom 1997). However, when 24-month-old children were exposed to multiple new words (eight in total), they performed well in the referent selection task but showed poor retention in a delayed recall task after just 5 min (Horst and Samuelson 2008). These findings suggest that while fast mapping may enable initial mapping under minimal exposure, its effectiveness diminishes when cognitive load increases, such as when multiple mappings are learned at once. To account for long-term retention, researchers have proposed “extended mapping”, a gradual refinement of memory traces through repeated exposure and use, as a critical mechanism (Kucker et al. 2015; Swingley 2010; see also Bion et al. 2013). Although repeated exposure plays a role, a central question regards how initial mapping differs between fast mapping, where novel meanings are inferred through contrast with familiar referents, and explicit instruction, which involves direct labeling without such contrast.

Addressing this question, a recent study found that although learning 10 novel word–meaning mappings through fast mapping resulted in lower initial mapping accuracy than with explicit instruction, it led to more stable memory consolidation after 1.5 years (Chen et al. 2023). This suggests that fast mapping may facilitate early integration of novel words into existing semantic and conceptual networks, supporting deeper encoding and more robust consolidation when novel items are contrasted with familiar objects (Atir-Sharon et al. 2015; Coutanche and Thompson-Schill 2014; Smith et al. 2014; Zaiser et al. 2022a, 2022b). Further evidence supporting this view has shown that fast mapping enables quicker semantic integration when the novel item shares multiple features with a familiar concept (e.g., Zaiser et al. 2022b), suggesting that such sharedness facilitates access to existing conceptual networks. Additionally, by the time infants begin to understand language at about the age of one, they have already accumulated rich observational experience with objects, actions, and phonological forms through babbling and imitation, further supporting their initial single-word acquisition (Macnamara 1972; see also Werker and Hensch 2015). Consequently, early exposure, or the contrast of previously formed semantic memory traces during word–referent mapping, appears not only to facilitate familiarity with sounds and forms but also to enable infants to rapidly associate new words with relevant concepts.

This human ability stands in contrast to standard artificial neural networks (ANNs), which, despite recent advances, generally require extensive training involving thousands of learning events and vast datasets to achieve stable internal information representations (e.g., Devlin et al. 2019; Lake et al. 2017). Numerous learning methods and neural network variants exist, with the most successful approach being the classic supervised backpropagation learning rule (Rumelhart et al. 1986) and its various adaptations. Although this algorithm in neural networks has proven highly efficient in language processing, such as speech recognition, and word and semantic processing (Chen et al. 2017; Kriegeskorte 2019; LeCun et al. 2015; Rogers et al. 2004), it is typically slow due to weight adjustments propagating through multiple deep layers and is criticized for its biological implausibility (Mazzoni et al. 1991; Reilly 1999). Furthermore, the human brain operates under neuroanatomical constraints that shape learning in ways that are often overlooked in artificial models, yet these constraints have been considered essential for simulating higher cognitive functions, such as language and symbolic learning and processing (Breakspear 2017; Pezzulo et al. 2013; Pulvermüller et al. 2021; van Albada et al. 2021).

2 Rapid symbolic learning in mind and brain

Since the discovery of fast mapping as a mechanism for rapid novel word acquisition, its neural underpinnings have received growing attention. Electroencephalography (EEG), with its millisecond-level temporal resolution, has revealed rapid neural changes following minimal exposure to novel word–referent pairings (de Diego Balaguer et al. 2007; Friedrich and Friederici 2011; Shtyrov 2011; Shtyrov et al. 2022; Torkildsen et al. 2008). For instance, by 6 months, infants exhibit distinct neural responses when mapping novel words to referent objects in a picture–word mismatch paradigm (Friedrich and Friederici 2011). The efficiency of this mapping seems to vary with vocabulary size in 20-month-old children, evidenced by neural activation differences in fast mapping (Torkildsen et al. 2008), suggesting that prior learning experiences shape fast mapping mechanisms. Similarly, an EEG study on artificial language acquisition demonstrated neural response changes within about 1 min of exposure, indicating rapid structural information extraction (de Diego Balaguer et al. 2007). Changes in brain responses due to rapid learning have also been consistently observed in adults. Several EEG studies have shown that novel phonological forms can be mapped onto neural representations after only brief exposure, with learning-related changes emerging in perisylvian language cortices (Kimppa et al. 2015; Partanen et al. 2017; Shtyrov et al. 2010). Extending this, other studies have examined the mapping of novel word forms onto their referent objects, revealing enhanced neural responses during word–referent association (Mestres-Missé et al. 2007; Shtyrov et al. 2021, 2022; Vasilyeva et al. 2019; for a review, see Leminen et al. 2023). In an EEG study, Shtyrov et al. (2022) had participants learn 20 unfamiliar auditory word forms paired with novel visual referents, each presented 10 times alongside familiar objects in a fast mapping condition. In contrast, an instructed explicit learning condition used additional unfamiliar word–referent pairs without familiar co-presentations. Fast mapping elicited stronger neural responses as early as ∼200 ms, whereas explicit learning was associated with later-occurring activation patterns. Notably, greater left-lateralization, particularly in language-related regions, has been documented in fast mapping scenarios compared to explicit learning (Shtyrov et al. 2021). These distinct activation patterns suggest that fast mapping engages early, automatic lexical-semantic processing and may rely on pre-existing conceptual representations.

Although substantial evidence has demonstrated neuroplastic changes underlying rapid memory trace formation during symbolic learning, the cortical regions involved in word–meaning mapping and their functional roles remain a matter of debate. Grasping the meaning of words and their relationship to the external world has been shown to rely on large-scale, distributed neural circuits spanning multiple cortical regions. These circuits encompass both domain-specific regions, including action and perceptual systems, and domain-general hubs that act as bridges between them (Binder and Desai 2011; Kuhnke et al. 2023; Pulvermüller 2013; Pulvermüller and Fadiga 2010). Importantly, these cortical regions are not uniformly engaged across all word types; rather, their activation patterns vary depending on the semantic content of the words. High-resolution functional magnetic resonance imaging (fMRI) studies have shown that different word types or symbols (action vs. object words, or animal vs. tool words) selectively activate visual and motor cortices, reflecting different aspects of their meaning (Buccino et al. 2005; Grisoni et al. 2021; Hauk et al. 2004; Meteyard 2012; Shebani et al. 2022; for work on patients, see also Dreyer et al. 2015). Behavioral studies showed motor interference effects during action word understanding or graspable object recognition (e.g., Buccino et al. 2017; Shebani and Pulvermüller 2013), further supporting the role of sensorimotor regions in word meaning processing. Importantly, these domain-specific activations are thought to interact with multimodal hub regions in frontal, temporal, and parietal cortices, functioning as convergence zones where different types of word meanings are equally processed (Binder et al. 2009; Binder and Desai 2011; Bookheimer 2002; Patterson et al. 2007; Pulvermüller 2013). Although it is clear that various cortical regions are involved in word meaning processing, debate persists over the exact role these regions play in facilitating rapid learning and symbolic processing. This debate is particularly relevant given that rapid plastic changes in neocortical regions in fast mapping conditions appear to occur independently of hippocampal learning processes (Atir-Sharon et al. 2015; Merhav et al. 2015). Lesions in semantic hubs, such as the anterior temporal lobe, impair rapid mapping, suggesting their crucial role in word learning-induced cortical plasticity (Atir-Sharon et al. 2015; Merhav et al. 2015; Sharon et al. 2011). The existence of this direct neocortical route challenges traditional views of the semantic memory system, which posit that it is first supported by the hippocampus and only later by slow, gradual consolidation of neural representations in the neocortex (Davis and Gaskell 2009; Frankland and Bontempi 2005; McClelland et al. 1995). Although the anterior temporal lobe has been implicated in fast mapping, word retrieval (Merhav et al. 2015), and enhanced retention after overnight consolidation (Mohan et al. 2022), inconsistent findings leave its role controversial and not yet conclusive (Smith et al. 2014; Warren et al. 2016; for a discussion, see Cooper et al. 2019), questioning the reliability of the underlying mechanisms and the methodological approaches of previous studies conducted with adults. Sensorimotor regions have also been implicated in rapid word learning, with transcranial magnetic stimulation over motor areas shown to modulate neural activity during brief learning sessions (Vukovic et al. 2021). Despite advances in understanding the cortical dynamics and plasticity underlying rapid symbolic learning, the precise neural mechanisms underlying these processes remain unclear. In particular, it is yet not fully clear how the brain fine-tunes vast numbers of synapses across distributed cortical networks, comprising thousands to millions of interacting neurons, to form rapid and stable memory traces.

3 Computational modeling for symbolic acquisition

Mathematical brain models have emerged as powerful tools for exploring the neural principles underlying language processing and its emergence. This approach is particularly fruitful, as it allows detailed examination of how internal representations emerge at the cellular and synaptic levels during learning. Standard ANNs have successfully achieved great performance in language and other tasks, as well as replicated aspects of their disorders (Chen et al. 2017; Kietzmann et al. 2019; Kriegeskorte 2019; Lake and Murphy 2023; LeCun et al. 2015; Mitchell and Krakauer 2023; Rogers et al. 2004; Rogers and McClelland 2004), ranging from speech and object recognition to translation, word acquisition, and semantic processing. Despite their power and human-like abilities, they require thousands of examples to form stable internal representations, thus relying on a slow, gradual learning process, in contrast to how human brains acquire novel knowledge. Previous modeling efforts have attempted to capture the rapid learning principle through a two-stage semantic learning process, involving an initial stage where concepts are encoded, followed by a second learning phase of symbolic learning that builds upon the previously established framework (Clay et al. 2024; Mayor and Plunkett 2010; Plunkett et al. 1992; Regier 2005). Although these models provide valuable insights into semantic and rapid learning, they differ significantly from realistic neurophysiological and anatomical structures of the human brain and employ biologically implausible supervised training methods like backpropagation (Mazzoni et al. 1991; O’Reilly 1998). Consequently, their applicability is limited when it comes to understanding rapid, cellular-level information storage and identifying key cortical regions involved in symbol acquisition. For instance, Mayor and Plunkett’s (2010) model demonstrated rapid word–reference learning after minimal exposure using a simple two-layer model architecture that directly linked visual and acoustic representations (e.g., the referent and acoustic word form dog). However, this model did not incorporate the broader network of cortical regions involved in word meaning processing in the human brain (see Section 2). As noted above, it is widely agreed that the biological realism of neural models is crucial for exploring the mechanisms underlying higher cognitive processing, as the structural properties of the human brain play a fundamental role in shaping distinct human capabilities (Breakspear 2017; Pezzulo et al. 2013; Pulvermüller et al. 2021; van Albada et al. 2021).

Following this approach, spiking brain-constrained neural models have been developed to specifically mimic the anatomical and physiological properties of 12 cortical regions of the frontal, temporal, and occipital cortices, as well as their connectivity structures relevant to language and symbolic processing (see Figure 1B; Tomasello et al. 2018; and early modeling architectures by Garagnani et al. 2008; Wennekers 2007). This approach has yielded an initial understanding of how different symbols (action and object-related words; e.g., run vs. sun) are instantiated in the human brain (Carriere et al. 2024; Garagnani and Pulvermüller 2016; Tomasello et al. 2017; Tomasello et al. 2018). These models were equipped with learning mechanisms based on Hebbian plasticity (Hebb 1949) including long-term potentiation and depression, which are well established biologically (Artola and Singer 1993; Tsumoto 1992). Acquisition of object and action words was simulated by co-activating neurons in primary visual and motor areas, to mimic key features of real-world situations, in which word forms are semantically grounded in information about their related objects and actions, consistent with reports from developmental studies (Tomasello 2003; Vouloumanos and Werker 2009). As a result, the model showed the spontaneous formation of cortical circuits (or the so-called cell assemblies) working as functional units (Braitenberg 1978; Garagnani et al. 2008; Palm et al. 2014; Pulvermüller 2013; Wennekers 2007). These circuits were scattered across areas exhibiting richer connections, so-called “convergence zones” (Damasio and Damasio 1994) or “connector hub areas” (Van den Heuvel and Sporns 2013), for general semantic processing and to modality preferential areas implicated for the processing of specific semantic categories, that is, temporo-visual regions for object words and frontal-motor regions for action words (Figure 1A and C; Tomasello et al. 2017, 2018). Intriguingly, neural circuits recruited additional neurons in primary visual (V1) and temporo-occipital (TO) areas differently when the model was deprived of visual input (i.e., in blindness) at either early or late in acquisition (Tomasello et al. 2024, 2019b). This set of simulation results replicated and provided a mechanistic explanation for a range of neurocognitive studies showing the involvement of sensorimotor and hub regions in word meaning processing (Binder and Desai 2011; Buccino et al. 2017; Carota et al. 2024; Hauk et al. 2004; Pulvermüller 2013; Pulvermüller and Fadiga 2010; Shebani et al. 2022) as well as visual cortex recruitment in cases of blindness (e.g., Amedi et al. 2003, 2004; Bedny et al. 2011). The key biological principles playing a role included connectivity structure and correlation learning, while neural circuit expansion mechanisms further explained the recruitment of visual neural matter in blindness. Recently, these neural network models were employed to explore abstract word processing (Dobler et al. 2024; Henningsen-Schomers and Pulvermüller 2022), language-driven concept formation (Nguyen et al. 2024), and the emergence of verbal working memory (Carriere et al. 2024), offering neuromechanistic insights into key aspects of linguistic phenomena. However, these modeling efforts, like standard ANNs using backpropagation, still rely on extensive training to establish internal representations. Biologically plausible learning rules, such as Hebbian learning, require numerous repetitions of co-activated neurons for stable neural representations to emerge due to their incremental synaptic changes (Froemke et al. 2010; Hebb 1949). This raises the critical question of whether a mathematical implementation capable of achieving rapid learning, similar to that observed in the human brain, is feasible at all. One hypothesis suggests that fast word–meaning mapping relies on pre-existing knowledge acquired through prior experiences (Macnamara 1972; see also Werker and Hensch 2015; and also Section 1). But how can this be explained at the neural matter level?

Figure 1:

Schematic of the brain-constrained neural network and emergence of symbolic circuits. (A) Fronto-temporal-occipital areas relevant for language and semantic processing showing neural circuits emergence after word–meaning mapping. As a result of word learning, specific circuits of neurons distributed across the network were activated (one dot indicates one active model neuron; blue: object-word circuit; red: action-word circuit; yellow: both), shown in the black boxes representing areas. (B) Schematic display of the area structure and connectivity of the brain inspired model. (C) Bars show average numbers of neurons per area for object-word circuits (dark gray) and action-word circuits (light gray) averaged across multiple networks; whiskers show standard errors. Note the relatively stronger representation of object-word circuits in ventral-visual areas and that of action-word circuits in dorsolateral-frontal areas. Asterisks denote significant differences. (Figure adapted from Tomasello et al. 2017, 2018.)

Repeated interactions with the external world, through separate exposure to actions, objects, and word forms, activate specific neural populations in their respective cortical modalities motor, visual and lang´nge regions. According to Hebbian learning principles (Hebb 1949), this co-activation gradually leads to the formation of specialized neural circuits, distributed in distinct cortical brain regions (Braitenberg 1978; Palm 1982; Pulvermüller and Fadiga 2010; Pulvermüller et al. 2014; Wennekers et al. 2006). This process may enable efficient mapping between preformed neural circuits, such as those representing sensory concepts (e.g., visual features of an object) and those associated with linguistic representations (e.g., auditory and articulatory aspects of a word, see Pulvermüller 2023). For instance, a child frequently encountering different types of dogs would lead to the emergence of neural circuits across visual and temporal cortices that represent the concept of “dog”. Similarly, repeated exposure to the word dog, whether through hearing or babbling, leads to circuit formations in the perisylvian language cortex, reflecting auditory and articulatory features (see Figure 2A, left panel). In this scenario, the final step involves linking these preformed neural circuits through their co-activation when the child simultaneously encounters visual and linguistic input during symbolic learning (Figure 2A, right panel), thereby enabling rapid word–referent association.^[2]

Figure 2:

Rapid symbolic learning at the neural level and in a brain-constrained neural model. (A) Visual depiction of the pre-exposure phase and neural circuit formation for word form and visual object processing in separated instances, with rapid word–meaning mapping (symbolic learning) subsequently occurring post-exposure due to preformed circuits. (B) Brain-constrained neural network simulations depicting the pre-exposure phase and circuit formation within the network for object, action, and word form cell assemblies (CAs), with circuits primarily located in the visual, motor, and language cortices, respectively. Bar plots depict the formed neural circuit cell sizes across the different modeled regions, with asterisks indicating statistically significant differences. (C) Comparison of neural circuit formation across learning events in two scenarios: a two-stage learning model with pre-exposure to stimuli and a one-stage learning model without pre-exposure. The two-stage model demonstrates significantly faster mapping compared to the one-stage scenario. (Panels B and C adapted from Constant et al. 2023)

Recent simulation work confirmed the neural hypothesis for rapid word–referent mapping (Figure 2A). By employing a brain-constrained neural model equipped with Hebbian learning, the model simulated a two-stage word learning process by learning object and action words grounded in the sensorimotor system. First, the model was pre-exposed to phonological and conceptual information; this was followed by symbolic associative learning that enabled rapid linking of preformed representations to their corresponding meanings. Intriguingly, in the pre-exposure simulations, all neural circuits extended into central hub areas within the network (Figure 2B), including the anterior temporal lobe (AT), a region identified as crucial for fast mapping (Atir-Sharon et al. 2015; see also Zaiser et al. 2022a). This finding suggests that such hub regions might serve as the primary site for rapid linking, as all neural circuits for objects, actions, and word forms extended into those regions. As a result, the network demonstrated rapid linking in the symbolic learning phase, with successful word-related circuit mapping occurring after just two to three exposures and most circuits fully established within 13 presentations in the model (Constant et al. 2023). Moreover, comparing a pre-exposed model with one that was not (mimicking direct associative learning) showed neural circuits emergence substantially later, requiring over 50 learning events (Figure 2C). Thus, this suggests that Hebbian learning is able to show rapid learning, especially when preformed neural representations are present, aligning with fast mapping theories (Atir-Sharon et al. 2015; Mayor and Plunkett 2010). While word–meaning mapping occurred rapidly, neural circuits became more stable and strongly mapped with repeated word–referent encounters, leading to more robust representations (Kucker et al. 2015), and potentially also explaining poorer initial retention shown in previous behavioral studies (e.g., Horst and Samuelson 2008). Notably, robust representations emerged much earlier in models with prior exposure compared to those without, underscoring the critical role of pre-exposure in facilitating rapid and stable learning.

Although the model was able to offer some key aspects of the underlying brain mechanism, much work is still needed. For instance, the pre-exposure of symbols and referents required up to 3,000 presentations to reach stable internal representations prior to associative learning, still relying on slow pre-acquired knowledge acquisition, which raises the question of how much pre-exposure is needed in this regard. Furthermore, object and action words were trained in the same manner, without accounting for their specific semantic features and properties. For example, the temporal and contextual dynamics involved in learning the word kicking, such as observing someone actively kicking a ball, differ significantly from learning the word dog, which often involves prolonged visual perception. Additionally, actions are often grounded in more contextual scenarios and typically do not occur in isolation. Future simulations should explore these different semantic features in action- versus object-word learning, potentially providing a neural explanation for the long-standing debate about the delayed acquisition of verbs (typically depicting actions) compared to nouns (object categories; Nelson 1973; for a discussion, see Waxman et al. 2013). Also in this regard, the model’s input structure did not consider shared semantic features between similar concepts (“dog” vs. “cat”) that might be represented by shared neural material in the cortex (Martin 2007; Pulvermüller et al. 2014). This is especially important, as previous studies have shown that highly shared features between familiar and novel referents can significantly facilitate rapid learning, as overlap in semantic characteristics helps the brain efficiently map new information onto existing knowledge (e.g., Himmer et al. 2017; Mak 2019; Zaiser et al. 2022a).

4 Conclusion and desiderata for artificial neural networks for rapid symbolic learning

Computational models have made significant progress in exhibiting human-like abilities across diverse domains. However, most models still lack the extraordinary ability of the human brain to acquire knowledge rapidly. While certain computational approaches have successfully achieved rapid symbolic learning, brain-constrained neural networks, designed to closely replicate the structure and functionality of the human brain, have been able to provide neuromechanistic insights at both cellular and cortical levels. Nevertheless, further research is essential to fully grasp the neural mechanisms underpinning this remarkable cognitive ability. As outlined above, key questions include the extent of pre-exposure to verbal and conceptual information, the role of shared semantic features, and the influence of existing vocabulary size. Each of these factors could impact rapid synaptic changes and the functional roles of cortical regions differently. Further work could also explore human-specific brain structures that may enable human-specific rapid learning capacities. Previous studies have shown that the connectivity within the perisylvian language cortices, particularly the arcuate fasciculus, is significantly more developed in humans than in nonhuman primates (e.g., Rilling 2014). Prior simulations (Carriere et al. 2024; Schomers et al. 2017) demonstrate that constraining networks with human-specific connectivity structures plays a pivotal role in the emergence of verbal working memory. Another important aspect not yet considered is the role of laminar structure and differentiation of distinct layers in the cortex, which has been proposed to vary between primary and higher-order cortical regions (Hilgetag et al. 2022), an important aspect that needs to be tackled in future studies regarding their role in the rapid formation of symbolic representations.

A breakthrough in understanding not only advances theoretical knowledge but can also guide modelers in developing more efficient learning strategies for neural network architectures. This effort is particularly critical nowadays, as high energy consumption remains a major criticism of current models, which require substantial computational power to achieve optimal performance (Cai et al. 2017; Tabbakh et al. 2024; Thompson et al. 2020). This energy-intensive demand necessitates extensive water usage for cooling data centers, further exacerbating the environmental impact. One of the most effective approaches to improving efficiency is parameter reduction, which involves decreasing the number of trainable weights and biases in neural networks (e.g., Frankle 2018). By reducing the model’s parameter count, it is possible to lower computational requirements and energy consumption, making neural networks more sustainable and suitable for deployment in resource-constrained environments. However, insights from neuroscience and language research could also be highly relevant. A valuable insight that could already inform current models is the importance of pre-exposure, driven by linguistic and cognitive theories, alongside designing architectures that closely mirror cortical structures involved in language processing, for example, the role of connector hub regions in linking distributed neural circuits and their connectivity patterns. Other factors may be critical as well, as mentioned above, which need to be examined first. All in all, examining critical biological principles with brain-constrained neural networks could further clarify which of these are most relevant for rapid learning.

A significant desideratum is that most current modeling approaches rely on web-based text or image processing, limiting their ability to accurately capture life-like language acquisition (for discussion, see Lake and Murphy 2023). Although efforts to integrate discourse-level information exist, such as communication patterns from forum discussions (see, e.g., Johns 2021), word meaning acquisition is inherently multimodal (e.g., Holler and Levinson 2019; Karadöller et al. 2024). Linguists and neuroscientists have long emphasized that word learning is fundamentally social (Tomasello 2000), unfolding within interactions where language functions as a tool for communication (Austin 1975; Grice 1957; Searle 1969; Wittgenstein 1953). In real-life language development, children receive a combination of verbal and nonverbal cues about referents and the surrounding communicative context. For example, children learn novel words like salt through varied interactions and modalities that may convey distinct communicative purposes: naming (This is salt) links words to objects, while questions (Is the salt on the table?) and requests (Pass me the salt) embed word meanings in contexts aimed at achieving specific goals. Prosodic contours (Villani et al. 2025) and accompanying gestures or facial expressions (Kelly 2006; Nota et al. 2021) further contribute to how such utterances are interpreted, offering crucial multimodal input during word learning. Yet another critical pragmatic feature in novel word–meaning acquisition is the “action sequence structure” associated with their function in communication: a request is usually followed by grasping and handing over an object, whereas naming is often followed by looking at the object and signaling recognition (Alston 1964; Fritz 2013). These action sequences appear to be a critical feature for understanding communicative functions (Pulvermüller 2018). Neuropragmatic research supports this view, showing that various speech acts conveyed through identical linguistic utterances across various modalities (e.g., prosodic, gestural) rely on distinct neural networks (for a review, see Tomasello 2023; Tomasello et al. 2025; Pulvermüller 2016). For instance, request and question elicit very early motor cortex activation during understanding, as compared to naming or statement functions expressed with the same utterance (Boux et al. 2021; Egorova et al. 2014, 2016; Tomasello et al. 2019a, 2022; van Ackeren et al. 2016; see Figure 5 in Tomasello 2023). This early motor cortex activation likely reflects the typical action sequences these communicative acts trigger, such as reaching to grasp an object or giving a verbal answer (Alston 1964; Fritz 2013; Levinson 1983), and may have an impact on the rate of novel rapid word–meaning mapping across different communicative contexts and the formation of their underlying neural circuits. Standard ANNs, are in general limited in capturing such rich, pragmatic, interactional dimensions and multimodality inherent in human language learning. Addressing these challenges at leastin part requires neural architectures capable of processing simultaneous multimodal inputs (visual, auditory, motor) to account for crucial pragmatic features of communicative functions, as well as, other highly relevant communicative features during conversation (e.g., Tomasello et al. 2025). Brain-constrained neural models designed for multimodal integration may offer a promising first step in this direction.