In partnership with
Vrije Universiteit Brussel and the University of Namur described a study of agent communication on October 14. The research examines how populations of artificial agents can settle on shared labels through local exchanges. The paper appeared in Transactions of the Association for Computational Linguistics in August 2026. The universities' announcement is current; the underlying study is not newly published.
The experiment treats communication as a reference task. One agent must identify a target entity within a set and send a word-like signal to another agent. The receiving agent selects an entity, then both receive feedback about success or failure. Their internal vocabularies change through these interactions. No central controller supplies a shared dictionary or dictates the final conventions.
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
The signals are not ordinary English words. The researchers generate short, artificial forms and connect each form to a concept built from measurable features. Depending on the dataset, those features describe values such as image characteristics, wine measurements, mushroom categories, or exoplanet properties. This allows investigators to inspect which feature patterns a signal represents. It does not mean people can understand the agents' messages without examining those representations.
The evaluation spans 37 public datasets, including scientific measurements and image collections. In the standard setup, ten agents complete one million training interactions, then face 100,000 test interactions. The reported results average ten independent runs. The researchers tuned shared settings on one visual dataset and used them across the collection without further tuning. That design tests transfer across varied data, though all evaluations remain within the study's defined task.
1,000+ Claude Prompts Top Professionals Actually Use at Work
Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.
These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.
Sign up for Superhuman AI and get:
1,000+ ready-to-use Claude prompts to get real work done in minutes — researched, tested, and used by professionals at Google, Microsoft, and NASA
Superhuman AI newsletter (4 min daily) so you keep learning new AI tools and skills to stay ahead in your career — the prompts are just the beginning
A separate population-size experiment varies the group from two to 100 agents on the CLEVR visual scenario. It also tests longer training schedules, reaching ten million interactions. This tests scaling within one experimental setting, not whether unrestricted numbers of real systems will coordinate.
Results differ by dataset. Across seven featured scenarios, test-set communication success ranges from 95.77 to 99.76 percent. The paper defines success by whether the listener selects the target. Its separate conventionality measure tracks whether agents would use the same signal under matching conditions. That measure ranges from 86.55 to 96.11 percent in those scenarios. Neither measure establishes open-ended reasoning, human-level language, or reliable behavior outside the experiment.
The study also examines differences in how agents perceive data. Some experiments introduce perceptual variation, altered agent structures, sensor noise, or changes in the environment. The authors report that conventions can remain effective or adapt under these test conditions. These are experimental manipulations, not trials with deployed robots, vehicles, or industrial sensors. For image inputs, the paper uses features from a pretrained vision encoder, another step between raw pixels and agent communication.
That boundary matters because the university release discusses possible relevance to robots and other systems with different sensors. The paper does not report a field deployment or compare the method with a production communication protocol. It studies whether shared conventions emerge in controlled reference games. Moving from dataset experiments to physical systems would require tests of timing, safety, failure recovery, and interpretation across machines.
The contribution is a studied mechanism for forming compact, shared signals without a centrally assigned vocabulary. It gives researchers a way to examine how local feedback shapes group communication. The most useful next evidence would test those signals in longer tasks and on embodied systems, then measure how well people can inspect failures. For now, the results describe a specific experimental capability, not agents secretly developing human language.
Your SOC 2, handled end to end with AI agents.
Enterprise buyers will not put your product near their customer data without a SOC 2 report. Twenty-one state privacy laws now hold them accountable for the vendors they share data with, so their diligence lands on you.
Sprinto gets you audit ready in 14 days, across three working sessions. AI agents connect your stack, collect the evidence auditors ask for, and close the gaps as they appear. You approve, they execute.
Sprinto also answers the security questionnaires your buyers send, using the same evidence base, so security reviews stop holding up deals.
No compliance hire. No consultant. Your auditor signs off.
