Humans Learn Language from Situated Communicative Interactions. What about Machines?

Human languages are evolutionary systems that continuously adapt to changes in the communicative needs and environment of their users (Schleicher 1869; Darwin 1871; Maynard Smith and Szathmáry 1999). They emerge and evolve as a result of meaningful and intentional communicative interactions between members of a linguistic community. When children acquire their native languages as new members of such a community, they build up their linguistic knowledge by actively taking part in communicative interactions with their caregivers. They thereby face two challenges that can be considered fundamental to the language acquisition process. On the one hand, children need to work out the communicative intents of their interlocutors, effectively making sense of the utterances they perceive. On the other hand, based on these utterances and their “reconstructed” meanings, they need to be able to bootstrap a linguistic system that enables them to understand and produce utterances that they have never observed before.

The cognitive processes involved in the acquisition of language from situated communicative interactions have been extensively studied in the fields of developmental psychology and cognitive linguistics, where they form the basis of constructivist and usage-based theories of language acquisition (Piaget 1923; Boden 1978; Bruner 1983; Langacker 1987; Nelson 1998; Croft 1991; Givón 1995; Clark 1996; Tomasello 2003; Goldberg 2006; Bybee 2010; Lieven 2014; MacWhinney 2014; Diessel 2017; Behrens 2021). The ability of children to acquire language is thereby attributed to two broad sets of skills, for which Tomasello (2003, pages 3–4) coined the terms intention reading and pattern finding. Intention reading refers to the cognitive abilities that are concerned with the functional, meaningful dimension of linguistic communication. These include in particular the capacity of children to share attention, to follow and direct the attention of others through non-linguistic gestures like pointing, and to recognize the communicative intents of their interlocutors. Pattern finding refers to the cognitive abilities that enable language users to generalize across different communicative interactions. These concern in particular the capacity to recognize similarities and differences in sensory-motor experiences, and to use this capacity for perceptual and conceptual categorization, schema formation, frequency-based distributional reasoning, and analogical thinking.

Intention reading and pattern finding are complementary yet highly interdependent skills. Intention reading enables a language user to reconstruct the intended meaning of an observed utterance. It implements an abductive reasoning process by which the language user constructs a hypothesis about the communicative intents of their interlocutor. This hypothesis is constructed based on clues that are provided by the utterance on the one hand, and its situational embedding in a communicative environment on the other. Pattern finding then provides the ability to generalize over observed utterances and their reconstructed meanings. It implements an inductive reasoning process, which, over time, yields an inventory of productive schemata that constitute the linguistic knowledge of a language user. As small children start out without any prior linguistic knowledge, their intention reading process can only rely on environmental clues. For example, during a communicative interaction with a caregiver, a young child might hypothesize that the observed utterance bear-gone refers to the observed disappearance from sight of their favorite stuffed animal. This association can be made based on sensory experience and pre-linguistic reasoning only, without any prior knowledge about the compositional nature of the utterance. An association between an observed form and its hypothesized meaning that is reconstructed based on environmental clues only is referred to as a holophrastic construction. When the same child later hears the utterance ball-gone in the context of a certain spherical toy disappearing from sight, their pattern finding ability might enable them to infer the compositional structure of the observed utterance. Indeed, based on the previously acquired holophrastic construction that associates bear-gone to their favorite stuffed animal disappearing from sight, and the observation of the utterance ball-gone in the context of a certain spherical toy disappearing from sight, the child might infer through syntactico-semantic generalization that the form bear is associated with their favorite stuffed animal, that the form ball is associated with a certain spherical toy, and that the form X-gone is associated with the referent of X disappearing from sight. The associations of bear and ball with their respective meanings are called holistic constructions. Constructions that contain abstract slots, such as the association between X-gone and its meaning of X’s referent disappearing from sight, are referred to as item-based constructions. The constructions that emerge from the pattern finding process can in turn provide clues for the process of intention reading during future communicative interactions. For example, if the child later observes the utterance gimme-bear in the context of a request by a caregiver to hand over their favorite stuffed animal, they can recognize the form bear, relate it to its meaning, and hypothesize that gimme-X is associated to a request by their caregiver to hand over the referent of X. The interaction between the processes of intention reading and pattern finding is thus bidirectional. Pattern finding relies on intention reading for providing hypotheses about the intended meaning of observed utterances, while intention reading relies on constructions that result from pattern finding to constrain and navigate the space of possible meaning hypotheses. The inter-dependency between the processes of intention reading and pattern finding is schematically depicted in Figure 1, using the examples introduced above.

The abductive nature of the intention reading process entails that the meaning that a language learner attributes to an observed utterance has the epistemological status of a hypothesis. This hypothesis is likely and plausible given the linguistic and non-linguistic knowledge of the language learner on the one hand, and the communicative environment in which the utterance was observed on the other. Hypotheses resulting from intention reading are thereby not free from uncertainty or doubt, and might in some cases not even transfer to any other communicative situation. An example of this uncertainty, rooted in anecdotal evidence, would be the case of a Flemish toddler (2Y2M) occasionally observing the utterance goed-weekend (lit. good-weekend—Eng. Enjoy your weekend) at the end of a conversation between their parents and caregivers when leaving their day-care center. At home, the toddler starts to reuse the phrase as a farewell greeting whenever they leave the room for a few minutes, for example to fetch another toy or garment. The hypothesis of the toddler about the meaning and communicative function of the phrase goed-weekend is only partially accurate. While they clearly have managed to reconstruct certain aspects of its meaning, in particular its social function as a farewell greeting, they did not capture the compositionality of the phrase and its meaning of wishing the addressee an enjoyable Saturday and Sunday.

The uncertainty introduced by the intention reading process directly percolates to the constructions that are formed, whether they are holophrastic mappings or the result of a syntactico-semantic generalization process. As a language learner uses these constructions across a multitude of communicative interactions that are situated in a variety of environments, they refine their hypotheses and gradually deepen the entrenchment of constructions that generalize better across situations. At the same time, constructions that are in conflict or competition with more generally applicable constructions, thereby hindering successful communication rather than supporting it, will gradually assume a different function or disappear altogether from the language learner’s inventory of constructions. The evolutionary dynamics of creating, strengthening, and weakening associations between linguistic forms and meanings, on any level of abstraction, based on their successful or unsuccessful use in communication, thereby provides a way to overcome the uncertainty that is inherent to the intention reading process.

As the language acquisition process of a child further unfolds, the linguistic clues that they provide as a speaker and discern as a listener gradually become more fine-grained and reliable. Their language understanding and production processes become less reliant on immediate environmental clues, making displaced communication possible and enabling language users to take part in conversations about abstract topics (Hockett and Hockett 1960). Such conversations rely to a large extent on the interpretation of linguistic clues with respect to the background knowledge and belief system of the individual interlocutors (Van Eecke et al. 2023). Yet, the linguistic and non-linguistic knowledge that is involved remains, on the most fundamental level, grounded in real-world experiences. “Democracy” and “eternity,” for example, are often cited as conversational topics at the more abstract end of the spectrum (see, e.g., Löhr 2022). The corresponding concepts have no directly observable referents as they capture social, cultural, and linguistic constructs. However, it does not take much to argue that without the grounding of these concepts in an individual’s lived experiences with volition, authority, permanence, perishability, and boredom, for instance, conversations about these topics would be hollow and soulless, if they would ever take place at all.

In sum, humans acquire their native languages by actively taking part in communicative interactions with other language users. These interactions are meaningful, intentional, and situationally embedded in a communicative environment. The situated and communicative nature of the interactions is essential to the language acquisition process, as language learners crucially depend on clues provided by the communicative environment to make sense of the utterances they perceive. Through the interdependent processes of situation-based intention reading and syntactico-semantic pattern finding, language learners gradually build up an inventory of productive schemata that capture, on any level of abstraction, the relationship between linguistic forms on the one hand, and their meanings and communicative functions on the other. Importantly, the language acquisition process is communicatively motivated, personal, usage-based, and constructivist, in the sense that the linguistic system of each person is individually built up and shaped by their past experiences and interactions. Through the evolutionary dynamics of creating, strengthening, and weakening associations between linguistic forms and meanings based on their successful or unsuccessful use in communication, language learners effectively manage to build up a productive linguistic system that is tied to their own physical and cognitive endowment yet compatible on a communicative level with the linguistic systems of other members of their community.

Over the last decade, important advances in neural machine learning and infrastructure (Sutskever, Vinyals, and Le 2014; Vaswani et al. 2017), combined with the availability of huge text corpora, have led to previously unimaginable progress in the field of natural language processing. The main catalyst in this process has been the advent of transformer-based large language models (LLMs), such as BERT (Devlin et al. 2019), GPT-3 (Brown et al. 2020), PaLM (Chowdhery et al. 2022), BLOOM (BigScience Workshop 2022), and LLaMA (Touvron et al. 2023). Definitive solutions to a wide variety of NLP subtasks that were previously considered immensely challenging, such as machine translation, speech recognition, conversational question answering, and text summarization, now seem well within reach—if not already considered achieved (for an overview, see, e.g., Lauriola, Lavelli, and Aiolli 2022). The extraordinary performance of these models on tasks that are traditionally considered hallmarks of human intelligence has naturally sparked the interest of researchers in other disciplines, including cognitive science, psychology, and linguistics. The focus of these interdisciplinary investigations is typically on comparing the performance of humans and LLMs on a variety of tasks, with the goal of assessing to what extent LLMs can serve as a model of language understanding and intelligent reasoning in humans. Methodologically, these studies more often than not adopt an approach called probing (Hupkes, Veldhoen, and Zuidema 2018), in which textual input prompts are designed to elicit the generation of output text that is taken to unveil the knowledge implicitly captured by the probed LLMs (Vulić et al. 2020). Probing as such does not reveal any explicit reasoning processes or strategies, but leaves it up to the interpretation process of the human experimenter to draw any conclusions about what a model must have “known” or “thought” (Shiffrin and Mitchell 2023). While probing is an effective methodology to systematically investigate the types of problems that LLMs can handle or struggle with—thereby sometimes revealing interesting “curious failures” (Shiffrin and Mitchell 2023)—there is some danger that the output-based interpretation of the inner workings of LLMs becomes a speculative and unfalsifiable endeavor. This risk can only be mitigated by a constant awareness of how these models are constructed, what they do, and how they do it (Shanahan 2024).

A major result obtained through probing studies is that despite the exceptional performance of LLMs on many NLP tasks, they seem to struggle with human-like logical and pragmatic inferencing (Jiang and de Marneffe 2021; Hong et al. 2023). In particular, it is often observed that LLMs fail to reason in a human-like manner about the knowledge they seem to capture (Choudhury, Rogers, and Augenstein 2022; Weissweiler et al. 2022; West et al. 2024). Although the precise nature of this knowledge is still a very active field of investigation, sometimes referred to as “BERTology” (Rogers, Kovaleva, and Rumshisky 2020), it has been convincingly argued that the reasons why large language models struggle so much with logical and pragmatic inferencing are to a large extent ascribable to the fundamental differences between how large language models are constructed and how human languages are acquired (Bender and Koller 2020; Trott et al. 2023).

Large language models are rooted in the tradition of distributional linguistics, as pioneered by Joos (1950), Harris (1954), and Firth (1957). Distributional linguistics offers a framework for analyzing language in terms of “the occurrence of parts [...] relative to other parts” (Harris 1954). These parts can be of any nature and can be chosen to correspond, for example, to phonemes, words, or dependency relations. Let us consider the example of words, given the focus of today’s NLP techniques on the word (or subword/character group) level. A word-level distributional analysis of a language describes that language solely in terms of the occurrence of words relative to each other. Words are thereby represented in terms of their frequency of collocation with all other words of the language as observed in corpora of language use. A word’s frequency of collocation is captured in the form of a high-dimensional vector of which each dimension corresponds to one other word of the language. These other words are referred to as the context words of a given target word. Frequency of collocation is defined in terms of the number of times a context word appears in the context window of a target word, by which the context window spans a chosen number of words to the left and to the right of the target word. The vectors representing words are thus the result of counting word co-occurrences in corpora of language use. Interestingly, words that exhibit similar distributions, indicated by a small angle between their vector representations, tend to have similar meanings. For example, the words salmon and cod would both frequently appear in the neighborhood of swim, water, and dinner, and only rarely co-occur with democracy and eternity. On the other hand, cod would be collocated more often than salmon with chips and less often with river. Their overall similar collocational behavior thereby indicates a high degree of semantic similarity, whereas their collocational differences are taken to indicate a semantic difference. The assumption that collocational behavior can serve as a proxy to meaning has become known as the distributional hypothesis, or perhaps more famously through the quote “You shall know a word by the company it keeps!” (Firth 1957). The elegance of distributional linguistics lies in the fact that linguistic forms can be analyzed independently from any external factors (Harris 1954). Word-level distributional analyses can indeed be constructed based on raw text only, without the need to include etymology, history, semantics, pragmatics, or grounding in (knowledge of) the world, for example. At the same time, aspects of those other factors can only be captured as a side-effect of their influence on the distribution of words with respect to each other (Hill, Reichart, and Korhonen 2015).

While the foundations of the field of distributional linguistics were laid out in the 1950s, and its routine application to specific NLP tasks such as word sense disambiguation (Schütze 1992,1998) and essay grading (Rehder et al. 1998) date back to the 1990s, the wide-spread adoption of distributional semantics as the standard meaning representation in the field of NLP is much more recent. Turney and Pantel (2010) are often credited with drawing the attention of the NLP community at large on the potential of using distributional meaning representations. A major turning point in this direction was the introduction of the famous CBOW and Skip-gram architectures for efficiently estimating high-quality word vectors, along with their implementation in the word2vec software package and the public availability of pre-trained word vectors (Mikolov et al. 2013a, b). In contrast to the high-dimensional, “sparse” vectors traditionally obtained through counting word co-occurrences, low-dimensional, “dense” vectors, called embeddings, were now estimated by a neural network in a word prediction task. While the method is somewhat different, the underlying idea and intuition remain the same: Words are represented as vectors that reflect their observed collocational behavior in corpora of language use. Interestingly, the dense vectors that are obtained through machine learning methods turn out to outperform traditional sparse vectors on virtually every NLP task they are used for (Baroni, Dinu, and Kruszewski 2014), for reasons that are still not fully understood today (Jurafsky and Martin 2024). The next major milestone concerns the use of transformer networks (Vaswani et al. 2017) to learn word representations via a masked language modeling task (Devlin et al. 2019). Such a task consists in filling in randomly selected words that are blanked in a given stretch of text. The major difference with the Skip-gram and CBOW-based methods is that the resulting word embeddings are “contextual” rather than “static.” As such, a target word is no longer represented by a single vector, but by a different vector for each different context in which it appears. The fact that words are now represented on the level of their textual occurrence rather than their dictionary entry makes for a more fine-grained modeling of their collocational behavior, and consequently leads to a drastically improved performance on downstream NLP tasks. For a more extensive overview of vector-based word representation methods, we refer the reader to the recent survey by Apidianaki (2022).

The scale on which today’s large language models capture the collocational behavior of words has led to a situation in which the texts that are generated do not only exhibit human-like lexical and morpho-syntactic correctness, fluency, and eloquence, but in which the resulting texts are also characterized by a remarkable discourse coherence and semantic adequacy. The contextual embeddings learned by these models mirror the cognitive semantic spaces of humans to a remarkable extent (Vulić et al. 2020) and even seem to capture non-linguistic knowledge to an extent that they can be used to solve natural language understanding tasks that were always thought to require vast knowledge of the world or human-like intelligent reasoning capabilities (cf. Hu et al. 2023; Webb, Holyoak, and Lu 2023). The reasoning errors made by LLMs under certain circumstances even resemble those observed in humans (Dasgupta et al. 2022).¹ Yet, these models still fully subscribe to the distributional linguistics paradigm, where languages are modeled solely in terms of the observed “co-occurrence of parts [...] relative to other parts [...] without intrusion of other features such as history or meaning” (Harris 1954). As such, all “knowledge” and “reasoning capabilities” they might hold are rooted in contextual word prediction based on collocations that were observed in huge amounts of raw textual data. While it is truly inconceivable that such results are obtained based on word distributions only, admittedly distilled from equally inconceivable amounts of texts, the fact that these LLMs are learned in complete absence of situationally grounded and intentional communicative interactions lies at the origin of a number of inherent limitations:

• Hallucinations. Large language models seamlessly mix fact and fiction in the output they produce, a phenomenon referred to by the term hallucination. Hallucination is a direct consequence of the generative nature of the models, by which they use collocation-based patterns and structures observed in input texts to generate output texts that exhibit similar patterns and structures. The epistemological status of the generated texts is thereby uniform and may consequently correspond to facts, beliefs, opinions, or fantasies in human terms (McKenna et al. 2023). Simply put, very probable word combinations might very well not correspond to factual truths or non-factual information contained in the training data. The use of the term “hallucination” to refer to generated fantasies is somewhat misleading, as hallucinations result from exactly the same generative process as any other output text. The use of the label is thus necessarily the result of a post-hoc interpretation by the human interpreter. One could as well argue that all output of LLMs is hallucinated, but that their hallucinations are remarkably often semantically adequate.

• Human-like logic and pragmatic reasoning. Large language models suffer from an apparent dissonance between the vast knowledge they seem to capture and their difficulty to use this knowledge to perform human-like logical and pragmatic reasoning (Choudhury, Rogers, and Augenstein 2022; Weissweiler et al. 2022; Mitchell and Krakauer 2023; Hong et al. 2023; West et al. 2024; Mitchell, Palmarini, and Moskvichev 2024). This limitation is ascribable to two main reasons. The first reason concerns the lack of situational grounding during the training process.² LLMs are trained to generate output texts that exhibit similar (sub)word distributions as those observed in input texts. This task, at which LLMs clearly excel, perfectly aligns with the standard assumption of generative machine learning that models should generate output that follows the same distributions as the input they are trained on. However, when the task shifts from generating texts that reflect the distribution of words in texts written by humans to human-like logic and pragmatic reasoning, the standard machine learning assumption is no longer upheld. Indeed, when the distribution of words in texts is used to make predictions about the distribution of objects and events in the world, it is no longer guaranteed that the input and target distributions coincide. The distribution of objects and events in the world might well be unlearnable from the distribution of words in texts. Moreover, it is extremely hard to gauge where these distributions coincide sufficiently to justify accurate predictions and where they are too different (Shichman et al. 2023). The second reason why LLMs struggle so much with human-like logic and pragmatic inferencing concerns the absence of communicative intent during the training process (cf. Bender and Koller 2020). Human linguistic communication is intentional, in the sense that speakers seek to achieve an effect in their interlocutors (Austin 1962). At the same time, language is an inferential coding system, by which is meant that a speaker’s intentions are not losslessly encoded in the utterances they produce, but need to be inferentially reconstructed by their interlocutors (Sperber and Wilson 1986). Pragmatic inferences are drawn based on these communicative intents, whereby the redundancy of information is particularly meaningful (Grice 1967; Kravtchenko and Demberg 2022). For example, if your co-worker mentions that the door was locked when they entered the lab in the morning, this presupposes (at least) that the door is usually not locked when they arrive. Or if you explicitly mention that your friend paid for their coffee, this presupposes that they normally don’t pay for their coffee. These presuppositions do not result from the information encoded in the utterances, or from general knowledge of the world, but from a reconstruction of the intention that motivated the speaker to convey this information that would otherwise have been redundant or irrelevant. While pragmatic inferences play a central role in human communication, they are notoriously difficult to draw for LLMs (Chang and Bergen 2024; Ruis et al. 2023). This should not be surprising given the fact that LLMs are trained on raw textual utterances in the absence of clues that might reveal the communicative intents that motivated their production. At the end of the day, the distributional hypothesis states that the distribution of words reflects their semantic similarities and differences, but remains silent about why particular utterances are produced in the first place.³

• Data-hungriness. Large language models are learned from raw textual data through a word prediction task. Consequently, all aspects of meaning, reasoning, and world knowledge that we expect LLMs to capture need to be learned indirectly through the intermediary of their effect on the distribution of words. In practice, this requires access to text corpora that consist of hundreds of billions of tokens (Chang and Bergen 2024). For example, the GPT-3 family of models was trained on 300 billion tokens sampled from a base corpus of more than 500 billion tokens (Brown et al. 2020) and the Chinchilla model was even trained on 1.4 trillion tokens (Hoffmann et al. 2022). These amounts of data are necessary to stretch the distributional hypothesis from making predictions about semantic similarity on the word level to generating output that exhibits aspects of human-like intelligence. The number of tokens that an LLM is exposed to is several orders of magnitude larger than the number of tokens a human is exposed to, estimated at about 60 million at the age of 5 and 200 million at the age of 20 (Frank 2023). It should be noted, however, that such a comparison is virtually meaningless, as the nature of their input, namely, texts vs. situated communicative interactions, is fundamentally different.

• Sensitivity to biases. Large language models mirror the biases that are present in the data they were trained on (Nissim, van Noord, and van der Goot 2020; Vallor 2024), concerning, for example, age, ethnicity, gender, religion, and sexual orientation (Navigli, Conia, and Ross 2023). These biases do not even originate from the unjustifiable extrapolation of correlations observed in the world, but from texts, written by humans, that at times literally contain stereotypes, hate speech, or fringe opinions. While a laudable idea in principle, the curation of training data is not a scalable solution given the difficulty of this task and the sheer amount of training data that is needed.

In sum, LLMs learn language from texts. They meticulously capture the collocational behavior of (sub)words in a fine-grained manner and on an inconceivable scale. All knowledge they seem to capture and reasoning abilities they seem to possess originate from observing the distribution of (sub)words with respect to each other in corpora of human-written texts. While their capabilities are astonishing, improving on almost any NLP/NLU task they are applied to and stretching the distributional hypothesis further than anyone would have imagined, the fact that they are learned from textual data and pass over the situated, communicative, and intentional aspects of human language use lies at the root of a number of inherent limitations. These concern in particular the uniform epistemological status of their output, their limited ability to perform logical and pragmatic reasoning, their data-hungriness, and their susceptibility to biases.

The text-internal prediction task through which LLMs learn their linguistic capabilities sharply contrasts with the situated communicative interactions through which humans acquire their native languages. As we have argued above, the fundamental differences between the learning processes of humans and LLMs in turn lead to fundamental differences in the kind of linguistic and non-linguistic knowledge that is built up, as well as in how this knowledge is used to engage in communicative and non-communicative actions. In no way does this argument refute the role of (distributional) statistical learning in human language acquisition (cf. Saffran, Aslin, and Newport 1996; Aslin 2017), which we consider, like Tomasello (2003) and Kuhl, Tsao, and Liu (2003) among others, to be an integral part of a human’s broader pattern finding ability (cf. Section 1). It does, however, imply that text-based statistical learning is insufficient on its own, and that the key towards more human-like language learning in machines lies in more faithfully modeling the circumstances under which human languages are acquired and used to communicate.

The goal of the present section is to zoom in on approaches that incorporate aspects of the meaningful, intentional, situated, communicative, and interactional nature of linguistic communication with the aim of operationalizing more human-like language learning in machines. We start by briefly considering recent work that remains firmly rooted within the core of the LLM paradigm, but extends it beyond a purely text-based prediction task (Section 3.1). In particular, we touch upon the integration of input from other modalities and the use of reinforcement learning to better align the output of LLMs with externally defined criteria such as human preferences. We then move on to discuss a line of work that takes a very different perspective on the matter, as it explicitly aims to study how artificial agents can acquire languages through the processes of situation-based intention reading and syntactico-semantic pattern finding during situated communicative interactions (Section 3.2). We discuss two concrete experiments that operationalize this paradigm and show how the linguistic systems that are acquired by the agents in such a setting are of a fundamentally different nature than those acquired by LLMs.

The argument that the key towards more human-like language learning in machines lies in more faithfully modeling the circumstances under which human languages are acquired and used to communicate minimally calls for letting go of a strict interpretation of the distributional hypothesis. In other terms, the idea that the “co-occurrence of parts [...] relative to other parts [...] without intrusion of other features such as history or meaning” (Harris 1954) suffices to capture all there is to human linguistic communication can no longer be upheld. When it comes to LLMs, this implies that the paradigm needs to be extended to integrate information beyond the collocational behavior of (sub)words as learned from raw text corpora through a text-internal prediction task. Two main approaches to extending LLMs beyond observed (sub)word distributions have already proven to be valuable additions to the paradigm. While the first approach focuses on broadening the input to the learning process to other modalities than text, the second approach concentrates on better aligning the output of LLMs with externally defined criteria such as human preferences or task-specific ground-truth accuracies.

The high-level idea behind multi-modal language models is that information originating from multiple modalities can be integrated into an LLM’s embedding space (see, e.g., Radford et al. 2021; Baevski et al. 2022; Lyu et al. 2023).⁴ The theoretical promise of this approach is that models capable of cross-modal inferencing break free from the limitation that knowledge of the world, or any knowledge other than word distributions per se, can only be captured as a side-effect of the influence of external factors on the observed distribution of words (see Section 2). For example, the graphical structures that are commonly used to represent spatial and temporal relations through the visual modality, for example through maps or timelines, concisely hold information about distances, durations, and temporal orderings that are hard, inefficient, or practically impossible to capture using the textual modality (see, e.g., Li et al. 2020). After all, there are good reasons why humans make extensive use of graphical representations in their atlases, road maps, and handbooks on subjects such as anatomy, botany, and chemistry. Apart from the theoretical advantages that come with enriching the representational power of LLMs beyond observed (sub)word distributions, the integration of non-textual modalities has also facilitated the development of a variety of real-world applications, such as multi-modal chatbots (OpenAI 2023; Yamazaki et al. 2023), robotic manipulation planning systems (Driess et al. 2023), and navigation aids (Fan et al. 2023).

It has often been argued that the text-internal training objective of LLMs yields models that might well generate texts that are in some sense indistinguishable from human-written texts, but at the same time fail to satisfy user expectations in terms of helpfulness, relevance, factual correctness, safety, and adherence to human values, among others (Weidinger et al. 2021; Liu et al. 2022; Tan et al. 2023). A particularly successful approach to better satisfying user expectations concerns the incorporation of text-external objectives through a two-stage language model alignment process. Concretely, a standard LLM is first pre-trained on a text-internal prediction task. Then, the learned model is fine-tuned to optimize a text-external criterion such as human preference or ground-truth accuracy on a specific task. The language model alignment approach has perhaps become best known through its use in combination with reinforcement from human feedback to align the output of open-domain chatbots to user expectations. Typically, a reward model is first trained in interaction with humans to capture how humans would rank a set of possible output texts. Then, this model is used as a reward function to fine-tune the pre-trained LLM. For a more detailed introduction into language model alignment, we refer the reader to Ziegler et al. (2019), Ouyang et al. (2022), and Bai et al. (2022).

The integration of input data from modalities other than text as well as the alignment of LLMs to text-external criteria cautiously lift the limitation that all knowledge, reasoning capabilities, and linguistic behavior of LLMs needs to be learned through the intermediary of observed (sub)word distributions. While the integration of input from multiple modalities provides a way to incorporate richer representations of the situational embedding of observed utterances, the alignment of LLMs to text-external criteria facilitates the integration of any optimization criterion for which an effective reward function can be designed or learned. Naturally, this sparks the question of whether the limitations ascribed in Section 2 to the absence of situationally grounded and intentional communicative interactions during the training process of LLMs still persist. Any answer to this question crucially hinges on two variables that were left open in the argument.

The first variable concerns the nature of the multi-modal input to the learning process, in particular the extent to which the input can reflect the sensory-motor interactions in which humans engage every day. There seems to be no reason to believe that it would be fundamentally impossible to collect training data in real-world situations using sensors modeled after the sensory apparatus of humans.⁵ At the same time, no comprehensive data sets of this kind exist to date and their construction would require overcoming extensive technical, ethical, and financial challenges. For now at least, the input to multi-modal LLMs typically remains restricted to a combination of texts, speech recordings, images, and/or videos.

The second variable concerns the reward function that is used in the language model alignment process. The central question here is to what extent such a function can capture human linguistic and non-linguistic behavior. The situation is much thornier than in the case of the first variable. The design of robust and reliable reward functions is notoriously difficult even in simple environments (Skalse et al. 2022; Ngo, Chan, and Mindermann 2024). The underlying reason is closely related to Goodhart’s law (Goodhart 1975; Manheim and Garrabrant 2018), which states that measures that become targets cease to be useful as measures (Ngo, Chan, and Mindermann 2024). For instance, the number of times the entrance door to a shop opens and closes might well be an excellent proxy to assess how busy that shop is. However, rewarding employees for maximizing this number is unlikely to be effective at making the shop any busier, as employees might spend all their time flipping switches and keeping clients out so that the door can be opened and closed more efficiently. If the goal is to capture something as multifaceted as human linguistic communication, it seems highly unlikely that an entirely non-gameable proxy can even exist. At the same time, it is difficult to give up on proxies altogether, as the alternative of replacing every invocation of the reward function by an authentic real-world interaction is definitely not feasible in such a data-hungry setting. Reward functions that rely on sufficiently realistic simulations of the situated communicative interactions that humans engage in might offer a way out, but they will somewhere need to strike a balance between robustness and versatility on the one hand, and computational efficiency on the other.

Do the inherent limitations of LLMs then persist when extending the paradigm beyond modeling (sub)word distributions? The argument developed in Section 2 that the tendency of LLMs to hallucinate, their limited ability to perform human-like logical and pragmatic reasoning, their data-hungriness, and their sensitivity to biases are limitations that are inherent to the paradigm was based on the premise that LLMs are learned from texts, optimizing a text-internal prediction objective. If the input shifts from raw texts to human-like sensory-motor observations and if the text-internal prediction objective is complemented with a human-like linguistic communication objective, the original argument no longer holds. The reality is, however, less clear-cut. While current LLM extensions beyond (sub)word distributions already exhibit more desirable behavior than text-internal LLMs, they are still far removed from the ideal, both on the input side and on the task side. The most fundamental question here seems whether a non-gameable proxy that is both practically usable for language model alignment and sufficiently close to capturing human linguistic behavior could in principle be designed or learned. In the end, the extent to which the limitations of LLMs can be lifted when extended beyond (sub)word distributions is bounded by the extent to which the meaningful, intentional, situated, communicative, and interactional aspects of human linguistic communication can be integrated into the training process.

The goal of the present section is to discuss an alternative approach that leaves the text-based prediction paradigm altogether and takes situated communicative interactions as the starting point for modeling language acquisition in machines. In particular, we examine a line of work that studies how artificial agents can acquire languages through the processes of situation-based intention reading and syntactico-semantic pattern finding during simulated communicative interactions. It is not our aim to provide a comprehensive overview of computational models of human language acquisition or emergent communication, for which we refer the reader to Doumen et al. (2025), Steels (2011), and Lazaridou and Baroni (2020), nor to provide an application-level alternative to current LLMs. Instead, we discuss two concrete experiments that show how the linguistic systems that are acquired by the agents in such a setting are of a fundamentally different nature than those acquired by LLMs. Through these experiments, we aim to further convince the reader that more faithfully modeling the circumstances under which human languages are acquired is likely to constitute a major leap towards more human-like language processing in machines.

The two experiments that we discuss below highlight different aspects of the approach. The first experiment, which is presented in Section 3.2.1, demonstrates how a population of autonomous agents can learn to communicate about entities that they observe in their environment. Concretely, by taking part in meaningful and intentional situated communicative interactions, the population converges on a communicatively adequate linguistic convention. Through the processes of intention reading and pattern finding, each individual agent builds up an inventory of holistic constructions that associate atomic forms to concept representations that are grounded in their own sensory endowment and experiences. The second experiment, presented in Section 3.2.2, focuses on the acquisition of constructions that capture compositional patterns of higher morpho-syntactic and semantic complexity. We consider a tutor-learner scenario in which the learner agent has previously acquired an inventory of conceptual distinctions that are grounded in the environment in which the agents operate, for example through the mechanisms presented in the first experiment. Through the processes of intention reading and pattern finding, the learner agent now bootstraps an inventory of item-based and holistic constructions that they can use to ask and answer English questions about their environment. The meaning side of these constructions is again grounded in the agent’s personal endowment and experiences.

The first experiment illustrates how a population of artificial agents that are equipped with sensors can establish a linguistic convention that is adequate to refer to entities that they perceive in their surroundings. The emergent convention consists of holistic constructions that associate atomic forms to concept representations that are grounded in the agents’ environment. Not only does this experiment show how the acquisition of situationally grounded concept representations through task-oriented communicative interactions can concretely be operationalized in artificial agents but it also shows how these concept representations can emerge and evolve to satisfy the communicative needs of a community of language users. The foundations of the methodological framework that is adopted date back to pioneering work by Steels (1995), Batali (1998), and Oliphant (1999), among others. The concrete experiment that we discuss was originally introduced by Botoko Ekila et al. (2024a, b). A skeleton version of the methodology is presented below to better fit the illustrative role of the experiment in this article.

In order to set up the grounded concept learning experiment, we first define a population, a world, and an interaction script. The population consists of a number of autonomous agents, say 10, which are each endowed with a sensory apparatus. The world consists of a number of entities that are represented through feature vectors. Each dimension of these vectors corresponds to a particular sensor that the agents are endowed with. Imagine, for example, that the world consists of geometrical objects and that the agents are endowed with 20 visual sensors. As such, the agents would perceive individual objects along 20 dimensions, say the number of corners of an object, its width-height ratio, its color channel values, its area, and its position on the horizontal and vertical axes, among others. At the beginning of the experiment, agents do not know any words or concepts. Any objects they might encounter in their environment are perceived as continuously valued feature vectors resulting from sensor read-outs.

The agents in the population engage in a series of situated communicative interactions that follow a structured interaction script. The script that is adopted in this experiment, and which will be formally defined below, proceeds as follows. First, a scene is created as a random subset of entities from the world. We will assume for now that a scene consists of a minimum of three and a maximum of ten entities. Then, two agents are drawn from the population and are randomly assigned the roles of speaker and listener for the purposes of the current interaction. One entity from the scene is randomly selected to be the topic of the interaction. As such, it will be the task of the speaker to draw the attention of the listener to this specific entity in the scene. Obviously, the identity of the topic is disclosed to the speaker only. The speaker then retrieves the construction in its construction inventory that best combines entrenchment and discriminative power for the topic with respect to other entities in the scene. If no adequate construction exists, as will often be the case in the initial stages of the experiment, the speaker invents a new construction that couples a concept representation, created based on the observed feature vector, to a new word form, say demoxu. The form of the selected construction is then uttered to the listener. If the listener already knows a construction of which the form side matches the utterance, they identify the entity in the scene that most closely matches the meaning side of the construction and point towards it. If the listener does not know a construction of which the form side matches the utterance, they indicate that they did not understand. In both cases, the speaker then provides feedback to the listener by signaling communicative success or failure and by pointing towards the topic entity. In case of success, both agents boost the entrenchment score of the constructions that they have used and inhibit the score of any competing constructions, that is, constructions that would also have discriminated the topic in the scene. At the same time, the concept representations of the used constructions are shifted towards the perceived feature vector. In case of failure, the speaker will inhibit the entrenchment score of their used construction. If the listener knew a construction of which the form side matches the observed utterance, they will inhibit its entrenchment score and shift its meaning side towards the observed feature vector. If they did not know such a construction, they will learn a new construction that couples the observed utterance with a concept representation that is created based on the perceived feature vector. A schematic overview of a situated communicative interaction of the grounded concept learning experiment is presented in Figure 2. The figure shows an interaction in which the speaker could successfully use the utterance walibu to draw the attention of the listener to the topic entity in the scene.

Formally, an experiment E = (W, P, G) is defined as a tuple that groups a world W = {e₁,…, e_m} comprising m entities, a population P = {a₁,…, a_k} of k agents, and a sequence $G=(gj)j=1i$ of i situated communicative interactions. Each agent is equipped with a set of l sensors S = {s₁,…, s_l} and is initialized with its own empty construction inventory I_a = {}. Each entity in the world is represented by means of a continuously valued l-dimensional feature vector X, of which the dimensions correspond to properties of the entities that can be recorded by the agents’ corresponding sensors.⁶ A holistic construction w = (f, c, s) is defined as a pairing between a form f ∈ F, a concept representation c, and an entrenchment score s. Entrenchment scores are bounded between a minimum of 0 and a maximum of 1. F is an infinite set of possible forms. Concept representations are each defined as a sequence of l tuples c = ((ω₁, μ₁, σ₁),…,(ω_l, μ_l, σ_l)) that group three numerical values: ω, μ, and σ. Each tuple relates to a feature channel of one sensor with which the agents are endowed. The weight value ω_i represents the importance of the feature channel for the concept, the mean value μ_i holds the prototypical value for the concept on this channel, and σ_i holds the standard deviation for the concept on this channel. Concepts are thus represented as a sequence of normal distributions. Each distribution relates to a feature channel and the relevance of the channel for the concept is indicated by a weight value.

The interaction script according to which each communicative interaction g ∈ G takes place is formalized as follows:

(A) Scene selection.n entities are drawn from the world to constitute a scene: C = {e₁,…, e_n}∈ W.

(B) Agent and role selection. A speaker S ∈ P and a listener L ∈ P are drawn from the population P.

(C) Topic selection. The topic entity T ∈ C is drawn from the scene C and disclosed to the speaker S.

If multiple discriminating constructions exist in the construction inventory of the speaker, the construction with the highest communicative adequacy is selected. The communicative adequacy reflects both the degree of entrenchment of the construction and the extent to which it discriminates the topic entity in the scene. The idea is that constructions that are more entrenched and better discriminate the topic are more likely to lead to communicative success. The communicative adequacy is concretely computed by multiplying the entrenchment score s of the construction with the difference in similarity between the construction’s concept representation c and the feature vector for the topic on the one hand, and the similarity between c and the feature vector for the closest other entity in the scene on the other. The form f of the construction with the highest communicative adequacy is then uttered by S as U and shared with L. If no discriminating constructions exist in the construction inventory of the speaker I_S, a new construction w = (f, c, s) is added, with f being randomly selected from the infinite set of forms F, s being assigned a default initial value, and c = ((ω₁, μ₁, σ₁),…,(ω_l, μ_l, σ_l)) being initialized with the exact values of the feature vector X for μ₁…μ_l, with a default initial value for σ₁…σ_l, and with a default initial value for ω₁…ω_l. The form f of this construction is then uttered as U and shared with L.

(E) Comprehension and interpretation. If a construction with the form U exists in the construction inventory of the listener, that is, w = (U, c, s) ∈ I_L, L points to the entity in the scene e ∈ C of which the feature vector X is most similar to c, according to the similarity metric defined in Equation (1). If no such construction exists in I_L, L signals that they could not understand U.

(F) Feedback.S signals success if L correctly identified T. Otherwise, S signals failure, and provides feedback to L by pointing to T.

If the communicative interaction g was not successful, S will decrease the entrenchment score s_S by a fixed value. If a construction with the form U existed in the construction inventory of L, that is, w = (U, c_L, s_L) ∈ I_L, L will decrease its entrenchment score s_L with a fixed value and shift its concept representation c_L towards X_T given C. If no such construction existed, a new construction w = (U, c, s) is added to I_L, with s being assigned a default initial value, and c = ((ω₁, μ₁, σ₁),…,(ω_l, μ_l, σ_l)) being initialized with the exact values of the topic’s feature vector X for μ₁…μ_l, with a default initial value for σ₁…σ_l, and with a default initial value for ω₁…ω_l.

The grounded concept learning methodology is directly applicable to any scenario where a population of agents needs to learn to communicate about entities that can be represented by means of continuously valued feature vectors. These vectors can originate from robotic simulations as referred to above or be derived from any tabular dataset that stores its entries in terms of continuously valued attributes. Botoko Ekila et al. (2024a) present three scenarios that show the broad applicability of the methodology on the one hand, and that illustrate the grounding of the agents’ emergent linguistic knowledge in the communicative task and situated environment on the other. In the first scenario, referred to as CLEVR, the scenes in which the interactions take place are based on the images from the CLEVR dataset (Johnson et al. 2017). Each synthesized 3D image depicts three to ten geometrical objects that vary in shape, size, color, shininess, and 3D position. The conversion from images to scenes is performed using the visual preprocessing procedure described by Nevens, Van Eecke, and Beuls (2020), resulting, per scene, in a set of 20-dimensional feature vectors that each represent one object in the original image. The 20 dimensions of the vectors correspond to human-interpretable visual properties, such as the object’s area, color channel values, position on the x and y axes, number of edges, and width-height ratio. A total of 1,000,000 training scenes and 100,000 test scenes were sampled from the original train and test splits, respectively. The scenes of the second scenario, referred to as WINE, were created based on the 4,898 physicochemical analyses of wine samples included in the Wine Quality dataset (Cortez et al. 2009). Each wine sample is described through an 11-dimensional feature vector in which each dimension corresponds to a particular physicochemical property of the sample, including for instance its acidity, amount of residual sugar, alcohol content, and amount of sulphates. Ninety percent of the wine samples were used to create 1,000,000 training scenes that each hold three to ten samples and the remaining 10% were used to create 100,000 test scenes in the same way. The scenes of the third scenario, referred to as CREDIT, were created based on the 284,807 financial transaction records included in the Credit Card Fraud Detection dataset (Dal Pozzolo et al. 2014). Each financial transaction is described along 28 dimensions that result from a principal component analysis. Again, 90% of the financial transaction records were used to create 1,000,000 training scenes that each hold three to ten records and the remaining 10% were used to create 100,000 test scenes. During the scene creation process, it was ensured that no scene holds the exact same entity more than once and that no entity that appears in a training scene is part of a test scene.

For all three scenarios, the parameters in the formal definition were set as follows. The population consists of 10 agents (k = 10) and each scene consists of a minimum of three and a maximum of 10 entities (⁠$3≤n≤10$⁠). Constructions initially receive an entrenchment score of 0.5. After a successful interaction, the scores of the used constructions are increased by 0.1 and the scores of their competitors are decreased by 0.02 * sim(c_q, c_r). After a failed interaction, the scores of the used constructions are decreased by 0.1. When it comes to the concept representations, initial channel weights (ω) are set to 0.5 and initial standard deviations (σ) to 0.01. Channel weights are rewarded or punished with a step of +1 and −5, respectively, on the sigmoid function $11+e−1/2x$⁠. Each experimental run consists of 1,000,000 communicative interactions, that is, one interaction for each scene in the training sets.

As the experimental runs unfold, the typical learning dynamics known from the language evolution literature (see, e.g., Blythe and Croft 2012) manifest themselves in all three scenarios. Figure 3 illustrates these dynamics for the case of the CLEVR experiment. The x axis represents the time dimension in terms of number of communicative interactions that have taken place. The solid line indicates, on the left y axis, the average degree of communicative success over the last 1,000 interactions. An interaction is counted as successful if the listener has indeed been able to identify the topic entity during the comprehension and interpretation step of the interaction script. Given that the construction inventories of the agents are empty at the beginning of the experiment, the degree of communicative success necessarily starts at 0. After 50,000 interactions, a degree of communicative success of about 90% is reached. This number then continues to increase, with the agents successfully communicating in 99.5% of the interactions after the 1,000,000 interactions of the experiment have taken place. The dashed line indicates, also on the left y axis, the average level of conventionality of the emergent language over the last 1,000 interactions. An interaction is counted as linguistically conventional if the listener would in principle have produced the same utterance if they would have been the speaker. Like the degree of communicative success, the level of conventionality will always start at 0. It increases at a somewhat slower pace than the degree of communicative success and reaches about 90% after 1,000,000 interactions. The dashdotted line shows, on the right y axis, the average number of distinct constructions used by the agents during the last 1,000 interactions in which they participated as the speaker. In the earlier stages of the experiment, a wide variety of concept representations and words emerges in the population, peaking at about an average of 90 constructions per agent after 50,000 interactions. As the emergent language becomes more conventional, the number of constructions in active use starts to decline, with an average of just over 50 constructions per agent after 1,000,000 interactions.

After running the experiments for 1,000,000 interactions, the emerged conventions are evaluated against the test portions of the datasets. The interaction script remains the same, with the additional constraints that agents can no longer create new constructions to add to their inventories and that construction scores and concept representations are no longer updated. The test results are reported in Table 1 for all three scenarios. Degrees of communicative success of over 99.5% are achieved in each scenario, validating that the emergent languages indeed generalize to previously unseen entities and scenes. Levels of conventionality range from 87.72% for CREDIT, over 88.67% for WINE, to 93.86% for CLEVR. These numbers show that the emergent languages are not only effective at solving the communicative tasks, but that the agents in the population also manage to effectively align their linguistic systems. The average construction inventory size of the agents ranges from 51.43 in the case of CLEVR to 52.67 in the case of WINE. While further analysis would be required to provide a more theoretically substantiated interpretation of the resulting construction inventory sizes, the reported numbers at least corroborate the finding that the agents converge on a manageable set of widely applicable concepts.

The emergent conventions consist of holistic constructions that associate linguistic forms with grounded concept representations. Figure 4 shows, for each scenario, a construction from the inventory of one agent that has reached the maximum entrenchment score of 1.0 after 1,000,000 interactions. The demoxu-cxn shown in Figure 4a emerged during the CLEVR experiment. It associates the form demoxu to a concept representation with three relevant dimensions, that is, dimensions with a strictly positive weight value ω. The area dimension represents the number of pixels within an object’s perimeter, normalized on a scale from 0 to 1. The bb-area dimension represents the normalized number of pixels within an object’s rectangular bounding box. Finally, the rel-area dimension represents the ratio between the number of pixels within an object’s perimeter and the total number of pixels in the image. Converting the normalized values back to raw pixel counts, the means and standard deviations on the three dimensions indicate that the concept representation of the demoxu-cxn prototypically represents entities with 1,344 pixels within their perimeter (σ = 76.8 pixels), 1,574 pixels within their rectangular bounding box (σ = 115 pixels), and which cover about 1% of the image. An analysis of the agent’s communicative behavior reveals that they use the construction 73% of the time to refer to small objects with a spherical shape. These are indeed objects that cover a relatively small area of the image and fill about 80% of their bounding box. The zapose-cxn shown in Figure 4b emerged during the WINE experiment. It associates the form zapose to a concept representation with a single relevant dimension, namely, the normalized residual sugar content of a wine sample. The prototypical value on this dimension corresponds to 12.34 g/l (σ = 1.39 g/l), a residual sugar content typically associated with medium dry wines. The ziseni-cxn shown in Figure 4c emerged during the CREDIT experiment and associates the form ziseni to a low value range on the second PCA component of the transaction records.

In sum, the grounded concept learning experiment shows how a population of agents that take part in situated communicative interactions can self-organize a linguistic convention that can be used to refer to arbitrary entities in their environment. The convention consists of holistic constructions that associate atomic forms to concept representations. The meaningful, intentional, interactional, and communicative nature of the learning environment facilitated the establishment of a linguistic convention that is not only grounded in the (non-textual) environment of the agents, but which is also motivated and shaped by their communicative needs. While we have focused on the emergence of such a convention, agents that are added to the population at a later stage can acquire the established convention using the exact same adoption and alignment mechanisms while they take part in communicative interactions with more “mature” agents. As such, also existing natural languages can be learned in a constructivist manner through interactions with “tutor” agents that already master these languages, as will be demonstrated in the second experiment.

Through the second experiment, we aim to illustrate how linguistic structures of a higher morpho-syntactic and semantic complexity can be acquired in a communicative and situationally grounded learning environment, thereby transcending the level of atomic word forms and concept representations addressed in the first experiment. Concretely, we focus on a tutor-learner scenario in which the learner agent needs to acquire in a constructivist manner a linguistic system that is adequate to ask and answer English questions about the environment in which the agents are situated. The experiment, which has been extensively discussed by Nevens et al. (2022) and Doumen, Beuls, and Van Eecke (2023), is set up as follows. The population consists of two agents, one being the tutor and the other being the learner. The communicative interactions between the two agents take place in randomly selected scenes from the CLEVR visual question answering dataset (Johnson et al. 2017). Each communicative interaction starts with the tutor asking an English question that is provided by the dataset for the selected scene. The task of the learner is to answer the question asked by the tutor. The learner starts without any linguistic knowledge apart from conceptual distinctions, such as different colors, materials, or sizes, which are assumed to have been acquired previously through a grounded concept learning experiment. The learner also possesses the ability to perform a number of primitive cognitive operations. These primitive operations can be thought of as the instruction set on which all complex reasoning processes need to build. In general, primitive operations can be cognitively inspired, rooted in theory of computation, or follow from a practical constraint such as a robot’s API specification. In our experiment, we provide access to a number of basic operations, such as segmenting a scene, counting the number of elements in a set, computing unions and intersections of sets, querying attributes, and filtering sets according to prototypes.

Imagine that during their very first communicative interaction, the learner observes the question How many blocks are there?. The learner cannot understand the question and receives the answer to the question, say 4, as feedback from the tutor. The learner then starts the process of intention reading. Based on the feedback provided by the tutor, the environment in which the utterance was observed, and the inventory of primitive cognitive operations that the agent can perform, the agent constructs a hypothesis about the intended meaning of the observed utterance. In order to do this, the agent searches for a network of primitive operations, referred to as a procedural semantic network (Woods 1968; Johnson-Laird 1977; Spranger et al. 2012; Verheyen et al. 2023), that, upon evaluation with respect to the scene, leads to the answer that was provided by the tutor. This network could, for example, consist of the operations [segment image → filter cube → count]. The mapping between the observed utterance How-many-blocks-are-there? and its hypothesized meaning [segment image → filter cube → count] is then stored by the learner agent as a holophrastic construction. Imagine that the learner later engages in a communicative interaction where they observe the utterance How many spheres are there?. Again, the learner cannot understand the question and needs to construct a hypothesis about the intended meaning of the utterance through the process of intention reading. Upon feedback by the tutor, in this case 3, the learner constructs the hypothesis [segment image → filter ball → count]. Based on the previously acquired holophrastic construction that associates the form How-many-blocks-are-there? with the meaning [segment image → filter cube → count] and the current observation of the utterance How-many-spheres-are-there? and its hypothesized meaning [segment image → filter ball → count], the learner can now use its pattern finding abilities to construct a more general item-based construction that pairs the form How-many-Xs-are-there? with the generalized meaning representation [segment image → filter ?type → count] in which X and ?type, respectively, represent variable slots on the form and meaning sides of the construction. Importantly, the construction also captures that the referent of the filler of the X-slot on the form side will fill the ?type slot on its meaning side. At the same time, the learner can acquire two holistic constructions that pair the forms block and sphere to their respective meanings, in this case the concepts of cube and ball. Along with the item-based construction and the two holistic constructions, the agent also learns that the block-cxn and the sphere-cxn provide suitable fillers for the X-slot in the how-many-Xs-are-there-cxn. This information is added to the agent’s categorial network, which models emergent grammatical categories as links between constructional slots and their observed fillers, very much in the spirit of Radical Construction Grammar (Croft 2001). Pattern finding is implemented as an anti-unification-based generalization process (Plotkin 1970; Van Eecke 2018; Yernaux and Vanhoof 2019). A schematic representation of the intention reading and pattern finding processes involved in this example is shown in Figure 5. Note that the constructions that are learned constitute form-meaning mappings that can be used for both language comprehension, that is, mapping from utterances to their meaning representations, and language production, that is, mapping from meaning representations to utterances that express them.

As discussed in Section 1 in the context of human language acquisition, intention reading is an abductive reasoning process that introduces uncertainty into the overall learning process. In this experiment, meaning hypotheses are guaranteed to be consistent and plausible given the communicative interaction at hand, but might not generalize to other situations. This uncertainty directly percolates to the holophrastic, item-based, and holistic constructions that are learned. The evolutionary dynamics that overcome this uncertainty are modeled on the level of constructions through similar alignment dynamics as those discussed in Section 3.2.1 in the context of the grounded concept learning experiment. Concretely, all constructions carry an entrenchment score, which reflects the frequency of their past successes and failures in communication. During language comprehension and production, constructions with a higher score will be preferred over less entrenched constructions. At the end of a successful communicative interaction, the learner agent increases the score of the constructions that were used and decreases the score of their competitors (i.e., other constructions that would have led to successful communication). In the case of a failed interaction, the scores of the constructions that were used are decreased. As a consequence of these evolutionary dynamics, constructions that are applicable in a wider range of situations gradually gain the upper hand over constructions that either compete with them or hurt communication rather than support it. Not only does this provide a way to overcome suboptimal hypotheses resulting from the intention reading process, it also causes more general and abstract constructions to ultimately prevail over more specific ones.

The learning dynamics of the experiment are visualized in Figure 6, where the degree of communicative success and the construction inventory size are plotted as a function of the number of communicative interactions that have taken place. The degree of communicative success starts at 0 and reaches 1 after about 25,000 interactions. The construction inventory size exhibits the typical “overshoot pattern” that was also found in the grounded concept learning experiment. In the earlier stages of the experiment, many new constructions are learned. These constructions are often holophrastic constructions or item-based constructions with few variable slots. Moreover, many constructions result from pattern finding operations over suboptimal meaning hypotheses. As more and more interactions take place in a variety of environments, the evolutionary dynamics of strengthening and weakening constructions based on their successful or unsuccessful use in communication ensures that suboptimal constructions gradually disappear from the construction inventory of the learner agent.

In sum, the grammar acquisition experiment shows how linguistic structures that transcend the level of atomic word forms and concept representations can be acquired through the processes of intention reading and pattern finding during situated communicative interactions. The linguistic knowledge that is built up consists in an inventory of constructions at different levels of abstraction. These form-meaning mappings correspond to syntactico-semantic generalizations over the compositional and non-compositional aspects of language use observed in meaningful, intentional, and situated environments. The meaning side of the acquired constructions is thereby grounded in the physio-cognitive endowment of the agents, their shared environment, and their communicative tasks and intentions. The evolutionary dynamics of strengthening and weakening constructions based on their successful or unsuccessful use in communication provides a way to overcome the uncertainty involved in working out the underlying meanings and intentions of linguistic utterances.

The two experiments discussed in this section illustrate that the linguistic systems built up by artificial agents during situated communicative interactions are fundamentally different from language models learned through a text prediction task. For one thing, the agents’ linguistic knowledge is directly grounded in their physio-cognitive endowment and environment. The constructions they acquire are motivated by their function in solving communicative tasks rather than by the observed collocational behavior of linguistic forms. For another, the processes of language comprehension and production no longer rely on (sub)word prediction, but on finding combinations of constructions that optimally map from linguistic forms to their meaning representations and vice versa. Due to their grounded nature, these meaning representations can be conceptualized and interpreted with respect to the communicative tasks and environment of the agents. As a consequence, the limitations of LLMs that result from their generative nature do not apply as such to the linguistic systems that are acquired in this setting. The agents do not hallucinate in the sense that LLMs do. They might well fail to achieve communicative success, but the utterances they produce are always motivated by more factors than the collocational behavior of the linguistic forms that they contain. The situated and interactional nature of the learning environment ensures that the constructions learned by the agents are communicatively motivated, which will ultimately be required to draw human-like logical and pragmatic inferences (see Section 2). The richness of the learning environment also lifts the constraint that all aspects of meaning, reasoning, and world knowledge need to be learned through the intermediary of their effect on the distribution of (sub)words, bearing the promise to avoid the need for exposure to hundreds of billions of tokens. Finally, the linguistic systems that are acquired will still mirror the biases present in the learning environment, but these will likely be easier to mitigate than those resulting from uncuratable amounts of textual data. By no means is this discussion intended to dismiss the impressive results obtained through the LLM paradigm, nor would we claim that achieving similar results through agent-based simulations is within reach in the near future. We do argue, however, that more faithfully modeling the situated, communicative, and interactional environments in which human languages are acquired provides a promising path to overcome the limitations of systems that essentially rely on the distributional hypothesis, as powerful as it might be.

The primary argument of this article has been that the way in which humans acquire their native languages is fundamentally different from the way in which LLMs learn their linguistic capabilities. We have argued that these differences have consequential repercussions on the linguistic knowledge that is built up, as well as on how this knowledge can be used to drive the processes of language comprehension and production.

We have highlighted the fact that humans learn language by actively taking part in meaningful and intentional communicative interactions that are situated in their everyday environment. During these communicative interactions, they make hypotheses about the meanings and communicative intentions that underlie the utterances they perceive, relying on clues provided by the communicative environment in which the interactions take place. Constructions of different degrees of abstraction are learned as syntactico-semantic generalizations over combinations of perceived utterances, their reconstructed meanings and intentions, and previously acquired constructions. As such, humans rely on their situation-based intention reading and syntactico-semantic pattern finding capabilities to bootstrap a productive linguistic system that is tied to their own physical and cognitive endowment, grounded in their environment, shaped by their past experiences, and motivated by their communicative needs.

By contrast, LLMs acquire their linguistic capabilities by learning to predict (sub)words based on the textual environment in which they appear. They meticulously capture the (sub)words’ collocational behavior in a fine-grained manner and on an inconceivable scale. Stretching the distributional hypothesis to previously unimaginable dimensions, they are able to generate texts that do not only exhibit human-like lexical and morpho-syntactic correctness, fluency, and eloquence, but that also exhibit a remarkable discourse coherence and semantic adequacy. The models even seem to be capable of solving natural language understanding tasks that require substantial non-linguistic reasoning abilities.

However, the fact that LLMs are learned in the absence of meaningful and intentional situated communicative interactions lies at the root of a number of inherent limitations. A first limitation concerns their so-called hallucinations. Due to their grounding in text generation, by which they use collocation-based patterns and structures observed in input texts to generate output texts that exhibit similar patterns and structures, they seamlessly mix fact and fiction in the output they produce. A second limitation concerns their struggle to perform human-like logic and pragmatic reasoning. For one thing, this struggle is ascribable to a lack of situational grounding during the training process, by which the distribution of objects and events in the world about which they need to reason differs from the distribution of (sub)words in the texts they are trained on. For another, it can be ascribed to the absence of communicative intent during the training process. While LLMs have been trained to capture fine-grained collocational patterns and structures in enormous corpora of language use, they have never had access to clues about why particular utterances were produced in the first place. A third limitation concerns their data-hungriness, which results from the fact that all aspects of meaning, reasoning, and world knowledge that they are expected to capture need to be learned indirectly through the intermediary of their effect on the distribution of (sub)words in corpora. Finally, LLMs are sensitive to biases, as they mirror the collocational patterns and structures present in uncuratable amounts of human-written texts.

Our comparison between the ways in which humans and LLMs learn their linguistic capabilities along with our analysis of the limitations inherent to LLMs have led us to conclude that the key towards more human-like language learning in machines lies in more faithfully modeling the meaningful, intentional, situated, communicative, and interactional aspects of human linguistic communication. In order to investigate how this could be operationalized in practice, we have first discussed approaches that extend the LLM paradigm beyond a purely text-based prediction task. The integration of input from other modalities than text provides a way to incorporate richer representations of the situational embedding of observed utterances, whereas the alignment of language models to text-external criteria confers the possibility of integrating task-oriented, communicative, and interactional objectives into the training process of LLMs. While these extensions have already proven to be valuable additions to the LLM paradigm, the extent to which situated communicative interactions can be approximated by non-gameable reward functions still remains an open question.

We have then moved on to discuss a line of work that leaves the text-based prediction paradigm altogether and takes situated communicative interactions as the starting point for modeling language acquisition in machines. We have discussed two experiments that model how artificial agents can acquire a language through the processes of intention reading and pattern finding during situated communicative interactions. The first experiment focused on grounded concept learning and shows how a population of autonomous agents can self-organize a linguistic convention that can be used to refer to arbitrary entities in their environment. The meaningful, intentional, interactional, and communicative nature of the learning environment facilitates the emergence and evolution of an environmentally grounded and communicatively motivated convention. The second experiment focused on the acquisition of linguistic structures that transcend the level of atomic forms and concept representations. It implements a tutor-learner scenario in which a learner agent acquires linguistic knowledge that is adequate to ask and answer English questions about their environment. This linguistic knowledge consists of constructions at different levels of abstraction, which can combine to map between English utterances and situationally grounded executable meaning representations. These meaning representations are grounded in the physio-cognitive endowment of the agents, their shared environment, and their communicative tasks and intentions.

The experiments have shown that the linguistic systems that are built up by the agents through situated communicative interactions are of a fundamentally different nature than the linguistic knowledge that is captured by LLMs. The agents’ acquired constructions are directly motivated by their communicative function, rather than by the collocational behavior of linguistic forms. The constructions can thereby be used to comprehend and produce linguistic expressions without relying on textual prediction, and the meanings of linguistic expressions can be conceptualized and interpreted with respect to the communicative tasks and environment of the agents. They thereby steer clear of the hallucination effects that characterize text prediction-based models. As a result of the rich nature of the learning environment, constructions can be acquired in a more data-efficient manner and they are able to capture the communicative function of language, which will ultimately be required to draw human-like logical and pragmatic inferences.

The experiments that we have discussed in the final part of this article definitely do not offer an application-level alternative to today’s LLMs, nor were they ever designed with that goal in mind. We hope, however, to have convinced the reader that a better integration of the situated, communicative, and interactional aspects of human linguistic communication constitutes the key to overcoming the limitations of current LLMs, and that more faithfully modeling the situated communicative interactions through which humans acquire their native languages provides a promising path towards more human-like language processing in machines.

We would like to thank Jérôme Botoko Ekila and Jens Nevens for their assistance in creating the graphs shown in Figures 3 and 6. We are grateful to Marie-Catherine de Marneffe and Remi van Trijp for their role in the discussions that led up to the writing of this article, and to Lara Verheyen, Remi van Trijp, and three anonymous reviewers for their insightful feedback on earlier versions of the manuscript. The research reported on in this article was funded by the European Union’s Horizon 2020 research and innovation programme under grant agreement no. 951846, the Flemish Government under the Onderzoeksprogramma Artificiële Intelligentie (AI) Vlaanderen programme, the AI Flagship project ARIAC by DigitalWallonia4.ai, and the F.R.S.-FNRS-FWO WEAVE project HERMES I under grant numbers T002724F (F.R.S.-FNRS) and G0AGU24N (FWO).