People who share no common language can still come to understand what a particular word refers to by repeatedly gesturing and pointing. Researchers at Belgium's Vrije Universiteit Brussel (VUB) and the University of Namur (UNamur) have reproduced this process of language formation in a population of AI agents. In an environment with no central coordinator and no shared dictionary, the agents were able to form a shared vocabulary grounded in observed data simply by repeating local conversations with one another. The findings were published in volume 14 of the peer-reviewed journal Transactions of the Association for Computational Linguistics (TACL), and the universities announced them on October 14, 2026. However, what was demonstrated is vocabulary formation in computer simulation; it does not mean that AI has acquired human-like language ability.
How AI without a shared dictionary forms its own vocabulary
The results are described in the paper "Emergent Communication in Continuous Worlds: Self-Organisation of Conceptually Grounded Vocabularies at Scale" (DOI: 10.1162/TACL.a.799). The authors are Jérôme Botoko Ekila, Lara Verheyen, Jens Nevens, Katrien Beuls, and Paul Van Eecke, researchers from the Artificial Intelligence Lab at VUB and the Faculty of Computer Science at UNamur. The paper was submitted in April 2026, revised in May, and passed peer review with Daniel Fried serving as action editor. It is published under a Creative Commons Attribution 4.0 (CC BY 4.0) license, and an earlier preprint (arXiv:2401.08461) is also available.
The research team tackled the question of whether a population of AI agents can acquire a shared language through dialogue, as humans do. Earlier research on emergent communication has also explored how agents create new symbols and words. But many studies dealt with simple environments with a limited set of features such as color and position, or with predefined categories, leaving open challenges in decentralized vocabulary formation based on complex observational data.
This study is distinctive in that it works with high-dimensional feature spaces reflecting real-world observational data and examines whether the whole population can form a shared vocabulary. The data include not only continuous numerical values but also discrete features representing categories. The team also tested cases in which each agent can observe different information, and cases in which sensor-equivalent inputs are lost midway. Combining multiple conditions, they examined whether shared conventions can emerge from interactions between individuals alone, without relying on a central coordinator or a shared dictionary.
In this setup, no shared vocabulary or concepts are distributed to the population. Each agent has its own independent internal state, and two individuals randomly selected from the population converse repeatedly. Depending on whether each conversation succeeds, each agent updates its own mapping between words and concepts. The design is such that, through the accumulation of these updates, a vocabulary that works across the entire population gradually forms.
Two agents converse repeatedly and acquire shared words
In the experimental environment, agents are not given existing categories such as "dog" or "car." The objects to be observed are represented as multidimensional feature vectors, and each agent perceives the characteristics of an object from those numbers. For example, an object is expressed as a combination of numerical values for color, shape, position, and so on, but how to classify it into a concept and which word to call it are not determined from the start. Through dialogue, agents form the concepts and words they need on their own.
The team prepared a mechanism for creating new words, a mechanism for representing concepts from observational data, and rules for updating both depending on whether a conversation succeeds. Concepts are represented using weights indicating the importance of each feature and parameters describing the distribution of observed values. These are combined into "concept prototypes," which are used to judge what kinds of objects a word applies to. Because each agent keeps its own word-to-concept mappings, there is no need for everyone to share the same internal parameters.
Conversations take place between two agents, a "Speaker" and a "Hearer." The Speaker chooses one object from several and utters a word to tell the Hearer which one it is. The Hearer compares the word it receives with its own concept prototypes and infers which object the Speaker is pointing to. If the Hearer identifies it correctly, the conversation is a success; if it gets it wrong, the correct answer is revealed, and the agents use it to update their word-concept mappings.
As this exchange is repeated, the words used to refer to the same object gradually converge across agents. If existing words cannot distinguish an object sufficiently, the Speaker can also create a new word. The Hearer, meanwhile, adjusts the scope of its concepts based on the outcomes of conversations and comes to understand the other's words. In this way, even though each individual learns from different experiences, a vocabulary that allows the population to communicate takes shape.
In the universities' announcement, Professor Katrien Beuls of the University of Namur stresses that there is no central computer deciding the rules. She explains that simply by conversing locally and reflecting the results, the individual agents cause conventions shared across the whole population to arise naturally. What matters is that the agents can correctly convey objects using the same words even if the concepts formed inside each agent do not match perfectly. What is shared is not the internal perceptual representation itself but the way symbols are used to communicate with others.
Tested on 37 datasets: success rates above 99% after one million interactions
To verify the effectiveness of the proposed method, the team used a total of 37 public datasets. These include physicochemical analysis values of wine, NASA data on exoplanets, mushroom characteristics, and features extracted from images. The image data include CLEVR, which deals with object attributes, as well as MS COCO, CelebA, and a bird image dataset. For real images, feature vectors were extracted with the pretrained image recognition model DINOv2 and used as input to the agents.
In the standard experiment, 10 agents carried out one million pairwise conversations, and the effectiveness of the resulting vocabulary was evaluated. Objects not used in training were also tested separately in 100,000 conversations. Results were aggregated from 10 independent runs, and in addition to communicative success, the team measured "convention agreement," which indicates how consistently agents use words, and the number of vocabulary items formed.
On three representative datasets, the communicative success rate on evaluation with unseen objects after one million training conversations was about 99.7%. The breakdown was 99.76% for CLEVR, 99.76% for the wine analysis data, and 99.67% for the exoplanet data. This means that agents starting without a shared dictionary came to convey objects correctly with high probability through dialogue. However, the same success rate was not obtained on every dataset.
For example, the success rate was 98.10% on the mushroom dataset, 95.77% on MS COCO with real images, 96.66% on CelebA, and 95.88% on the bird images. In the seven representative experiments, high success rates were obtained in every case, but results differed depending on the nature of the data. The figure of about 99.7% therefore cannot be applied to all 37 datasets.
The team also examined robustness when observation conditions differ between agents. The experiments included numerical offsets imitating sensor calibration errors, noise added to observed values, a mix of agents with different available sensor combinations, and situations where a sensor fails midway through a conversation. These were not produced by breaking physical sensors but by altering the input data.
As a result, the vocabulary system was maintained under many settings even when observation conditions differed. However, the effects were not negligible. For example, with strong noise added in CLEVR, the communicative success rate fell to 87.67%, and under conditions with larger differences in the sensors available to each agent it was 83.47%. Even when half of the sensor-equivalent inputs were lost midway, the success rate held at 93.82%, though variation between runs grew. The method shows a degree of adaptability, but the larger the differences in observation conditions, the greater the risk of performance degradation.
In the 10-agent CLEVR experiments, communicative success and the vocabulary formation process were evaluated over 10 independent runs, with the mean and a range of two standard deviations (±2σ) shown. The validation thus considers variation between runs rather than a single successful example. Still, all of these results come from computer simulation; they were not confirmed with real robots, self-driving cars, or industrial sensors in factories.
The advantage of building a language suited to sensors rather than relying on human words
The goal of this research is not to teach human words directly to AI agents but to have agents form a symbol system suited to the information they handle. Human language developed from human senses and experience, such as sight, hearing, and touch. But many sensors used by robots and autonomous systems capture information humans cannot perceive directly, such as infrared, LiDAR, sonar, and chemical analyzers. Insisting on translating such information into human words may fail to capture the features and distinctions needed for machine-to-machine communication.
Jérôme Botoko Ekila, a doctoral researcher at VUB, explains the significance of letting each AI system form its own concepts based on the information it observes, rather than forcing systems to conform to human language that is poorly suited to their sensors. Even for data that is hard for humans to understand, if agents can refer to it with shared symbols, it may help with information exchange.
Professor Paul Van Eecke of the same university explains the mechanism using the phrase "red car." Even among humans, the color values each person perceives when looking at the same red car do not match exactly. Yet if they share the words "red car," they can understand which car the other is pointing to, because language serves as an abstraction that absorbs differences in individual perception. The research team is trying to get AI agents to form such a mechanism autonomously.
This is also where the approach differs from large language models (LLMs). A typical text-centered LLM learns relationships between words and concepts from large amounts of text. In this study, by contrast, the mapping between words and meanings is formed from the observational data given to the agents themselves and from the success or failure of their conversations. The two differ in purpose and in the problems they address, so their performance cannot be simply compared, but there is a clear difference in how language is acquired.
| Comparison item | Typical text-centered LLM | Emergent communication method in this study |
|---|---|---|
| How language is acquired | Pretraining on large amounts of human-written text | Dialogue between agents and feedback on success or failure |
| Relationship between meaning and observational data | Mainly learns relationships between concepts from text | Directly links features of observational data to words |
| System structure | Centered on a large neural network | Many agents, each with its own internal state |
| Language updates | Model behavior adjusted through additional training, etc. | Individual vocabularies and concepts updated through dialogue |
| Interpretability of internal representations | Difficult to interpret representations inside the neural network directly | Word-to-feature correspondences are easy to trace |
| Differences in observation conditions | Handling varies depending on the model and training method | Assumes agents with different features and sensor configurations |
Note that some recent multimodal LLMs can directly handle information such as images and audio, so it is not the case that LLMs in general are disconnected from observational data. This comparison simply lays out the difference between typical text-centered learning methods and the decentralized vocabulary formation in this study. Rather than a technology to replace LLMs, the work is best positioned as research into how multiple agents with different sensory information can form a shared symbol system.
Another feature is that the meanings of the formed vocabulary are easy to trace. In this method, the concept corresponding to a word is represented using per-feature weights and probability distributions. Researchers can examine which features a word emphasizes and which observed objects it applies to. Furthermore, using dimensionality-reduction visualization, they can also see how concepts change as dialogue accumulates and how different words come to be used distinctly within the feature space.
The paper also shows how, as this differentiation of concepts progresses, each word comes to correspond to specific regions or objects. The research team analyzed this phenomenon as the formation of "niches" in conceptual space. The fact that word usage is adjusted through dialogue even when internal representations do not fully match supports the hypothesis that the conventionality and adaptability of language can arise from interactions between individuals. However, these results do not explain the evolution of human language as a whole, and should be taken as experimental evidence from a limited communication task.
Caution is needed with the phrase "AI acquired its own language"
VUB and UNamur position the study as an important result that could lead to communication between autonomous systems such as self-driving cars, robots, and industrial sensors. Indeed, if machines with different sensors can form shared symbols without a central coordinator, it could help distributed systems operate in coordination. However, these are future applications envisioned by the universities, not uses demonstrated in this study. What the paper carried out was computer simulation using public datasets.
For example, whether multiple vehicles can autonomously form a vocabulary using observations from the LiDAR and cameras mounted on self-driving cars was not tested in these experiments. In real environments, in addition to sensor degradation and changes in the surroundings, communication delays, packet loss, and sudden changes in observed objects also occur. Agents cannot always observe the same object, and it can be difficult in itself to confirm what another's utterance refers to. Whether a mechanism that works on numerical data remains stable under such conditions requires further verification.
Caution is also needed with the expression "AI has begun speaking its own language." What these agents acquired is a vocabulary system centered on "referring conventions" for pointing to specific objects. They have not become able to freely handle the complex grammar, tense, metaphor, or context-dependent implications seen in human language. The paper also tests utterances combining multiple words, but these too serve the limited purpose of correctly identifying objects. It would not be appropriate to interpret this as AI having acquired human-level language ability or intelligence.
The communicative success rate also needs to be understood in light of the experimental conditions. The high success rate of about 99.7% was obtained after one million training conversations on specific datasets, and success rates dropped under different observation conditions. Validation with 37 datasets is important for evaluating the method's generality, but it does not guarantee equivalent performance in every real-world environment. The results are at the stage of showing that decentralized vocabulary formation works under designed conditions and can adapt to a certain degree of change.
Challenges that remain before AI language formation reaches the real world
The key achievement of this study is showing that, in a decentralized environment with no central coordinator, multiple agents can form a shared vocabulary from observational data containing continuous features. In addition, the study examines several challenges within a single experimental framework: sharing a vocabulary across the whole population, handling different observation conditions, and updating concepts as the environment changes. What the agents share is not a finished dictionary but ways of using words that gradually take hold through interaction. This mechanism also offers a new clue for studying the "symbol grounding problem" of how the meanings of words become tied to observed objects.
One challenge ahead is scalability when the number of agents increases further. The team has already compared populations of 2, 10, 25, 50, and 100 agents using CLEVR, with validation of up to 10 million conversations. After one million conversations, the success rate was 99.76% with 10 agents but fell to 92.69% with 100 agents. On the other hand, even in the 100-agent population, increasing conversations to 10 million improved it to 99.16%. The result indicates that the larger the population, the fewer opportunities each agent has to take part in conversations, so more conversations are needed before a shared vocabulary forms.
Going forward, it will therefore be important to see whether vocabulary can be shared efficiently even in populations of hundreds or thousands of agents or more. In larger populations, unique words used only among some agents, or conventions equivalent to regional dialects, may arise. Further verification is also needed on whether the whole population can maintain mutual understanding when conversation frequency or the connections between agents change.
Another challenge is moving from static datasets to real physical environments. The experiments reproduce sensor failures and noise numerically, but with real machines, multiple unpredictable changes occur at the same time. Many points remain to be confirmed, such as whether robots equipped with sensors from different manufacturers can form a shared symbol system while absorbing discrepancies in observed values, and whether vocabulary sharing can be maintained even when communication is temporarily interrupted.
The full text of the paper can be read at the ACL Anthology, and the experimental conditions and evaluation results are also public. The arXiv preprint also includes an author note stating that generative AI tools were not used in conceiving or writing the paper. Going forward, it will be important to reproduce the results using this public information and to check whether the same properties are obtained in different environments.
The finding that AI agents can form shared symbols based on their own observational information without relying on human words opens new possibilities for research on distributed AI systems. However, high communicative success rates in simulation alone are not enough to judge practical usefulness. The next key focus will be whether machines with different perceptions can exchange information stably, as tested with real sensors and communication environments.