Give two agents a collection of objects represented by continuous measurements. Ask one to identify a target for the other. Do not provide a dictionary. Through repeated successes and failures, the agents can converge on labels of their own—shared words attached to learned regions of the data.[1]
Researchers from Vrije Universiteit Brussel and the University of Namur tested this decentralised method across 37 public datasets, including physico-chemical wine measurements, NASA exoplanet data and image collections of birds and faces. The resulting vocabularies were not copied from English. They were grounded in the features available to the agents’ sensors.[1][2]
A word is paired with a prototype
The system links a symbolic form—the invented label—to a subsymbolic concept representation in a continuous feature space. A prototype might occupy a region defined by several measurements at once. The label becomes useful when the speaker can use it to distinguish the target and the listener has learned a compatible meaning.[1]
As an illustrative example, imagine five wines described by acidity, sugar and alcohol. An agent might invent “mava” for a prototype near high acidity and low sugar. A successful reference reinforces the association; a failure prompts adjustment. This invented example explains the mechanism, rather than reporting a word from the experiment.
The paper supplies an actual example: in its exoplanet scenario, the form “penatu” became associated with terrestrial-type planets around 81% of Earth’s radius. Its meaning depended on both radius and a categorical planet-type feature. That is a concrete link between an arbitrary symbol and measurements the agents can compare.[1]
No agent owns the dictionary
The convention emerges through local interactions between pairs drawn from a population. There is no central authority assigning categories. Agreement accumulates because successful forms spread and incompatible ones lose ground. The study reports that the process remained robust under perceptual variation, including populations whose agents did not perceive the world in exactly the same way.[1]
That makes the result different from training a language model on human text. The agents are not learning to imitate existing sentences. They are solving a coordination problem: compress enough of a continuous environment into reusable symbols that another agent can act correctly.
On CLEVR, a dataset of rendered objects with controlled attributes, a 10-agent experiment averaged over ten runs rose from about 96% successful reference after 50,000 interactions to 99.81% after one million. Success meant that the listener selected the intended entity. It was a coordination score within this task, not a test of general language understanding.[1]
The vocabulary can change with the world
The authors report that the conventions self-adapt when the environment changes. That is important because a fixed codebook can become obsolete when sensors drift, objects change or new categories enter the task. An evolving convention can redraw a boundary or create a new label through the same interactions that built the first vocabulary.[1][2]
The prototypes also make the language more inspectable than an opaque message vector. Researchers can connect an invented word back to underlying feature values and ask what part of the sensor space it represents. That does not make every interaction self-explanatory, but it gives the vocabulary a traceable semantics.[2]
Why autonomous systems may need non-human words
A fleet of robots may share depth, vibration or chemical signals for which ordinary human vocabulary is clumsy. Industrial sensors may need a compact label for a pattern spread across dozens of measurements. A human-readable description can still be added, but the most efficient internal convention may begin with the machines’ own perceptual geometry.
The study therefore offers a middle path between a hand-written protocol and uninterpretable continuous messages. Symbols can emerge rather than be prescribed, while their meanings remain linked to measurable prototypes. The next challenge is not simply to make agents talk. It is to preserve that transparency as tasks become more complex, goals diverge and communication starts to influence the world rather than merely identify an item.
Sources & further reading
- Emergent Communication in Continuous Worlds: Self-Organisation of Conceptually Grounded Vocabularies at Scale Transactions of the Association for Computational Linguistics · Accessed 15 Oct 2026
- VUB and UNamur study shows how AI agents develop their own artificial natural language to communicate about complex data Vrije Universiteit Brussel · 14 Oct 2026 · Accessed 15 Oct 2026