AI agents build shared vocabularies from measurements without a central dictionary

Give two agents a collection of objects described by measurements and ask one to identify a target for the other. Without a shared dictionary, they must develop a convention through successful and unsuccessful exchanges. Researchers at Vrije Universiteit Brussel and the University of Namur have demonstrated a method for building those vocabularies across a wide range of datasets.[1][2]

The work appeared in Transactions of the Association for Computational Linguistics in August and was highlighted by VUB on 14 October. Jérôme Botoko Ekila and colleagues evaluated the method on 37 public datasets, including physical and chemical wine measurements, exoplanet data and image collections. Their question was how agents could connect shared symbols with the features each agent perceives.[1][2]

An invented word has a measurable meaning

Each word associates a symbolic label with a concept representation in feature space. The label itself can be arbitrary. Its usefulness comes from identifying a region of the measurements that distinguishes the intended object from alternatives. Successful exchanges reinforce the association; failures create pressure to adjust the word or its meaning.[2]

The paper gives a concrete example from its exoplanet scenario. The form ‘penatu’ became associated with terrestrial-type planets whose radius was around 81% of Earth's. Its meaning combined a continuous measurement, radius, with a categorical feature, planet type. Researchers could therefore inspect what the invented word referred to rather than treating it as an opaque message.[2]

Agreement grows through local interactions

Pairs of agents communicate within a population; no central authority assigns the vocabulary. Conventions develop as useful labels spread and incompatible ones lose ground. The mechanism allows the group to coordinate around categories suited to the task and the information available to its members.[1][2]

In a CLEVR experiment with 10 agents, communicative success rose from about 96% after 50,000 interactions to 99.81% after one million, averaged over ten runs. Success meant that the listener selected the intended entity. It is a measure of coordination within this referential task, rather than a score for general language understanding.[2]

The study also examined perceptual variation. Agents did not always receive identical measurements or access to the same features. The reported results show that useful conventions can emerge under such differences, although stronger noise and reduced shared perception make the task harder. That gives the method a more realistic test than requiring every agent to see an identical world.[2]

A vocabulary can change with its environment

The authors report that the learned conventions can adapt when the environment changes. A fixed codebook may become less useful when the distribution of objects shifts or a sensor behaves differently. Here, the same interaction process that produced a category can refine it as the task changes.[1][2]

For autonomous systems, the attraction is a compact way to communicate about measurements that may not fit ordinary human words. A group of machines could need a label for a combination of shape, temperature or other features. This is a potential application of the principle, not a deployment demonstrated by the paper.[1]

The work offers a useful combination: symbols emerge from interaction, while their meanings remain connected to inspectable measurements. It does not demonstrate human grammar or culture. Its contribution is grounded coordination—agents can develop a shared way to refer to their world without having every category specified in advance.[1][2]

Report an error or suggest a correction