When It Pays for Two AIs to Exchange Messages: The Study

Update, November 3, 2026: we removed references to agent networks and medical agents, which added no evidence to the study’s question, and the questions that repeated the text.

Having two AIs exchange messages is not enough to make them work better together, as a preprint (a study not yet peer reviewed) published on arXiv shows. The authors studied classifiers, meaning programs that assign an object to a category, each of which observes a different part of the same object and sends short messages to a receiver.¹ With two-bit messages designed jointly, the team gained on average between 18.07 and 22.68 percentage points over messages designed separately. However, it is not a given that two senders beat a single one using the same communication budget.¹

What experiment is this about?

The study’s starting point is that two machines can each recognize an object well on their own and still make mistakes when they have to work as a pair. In the setup described, each classifier sees only a portion of the object and communicates with a receiver, which has to guess its class.¹ The tests were run on three digit datasets, including MNIST. The results therefore hold for this type of task and not for AI systems in general.¹ Message length is measured in bits, the smallest unit of information. The shorter the message, the narrower the vocabulary available to the sender, and so how it is used matters.¹

What is the difference between coordination and cooperation?

The authors separate two questions. Coordination concerns how to design the messages together, so that senders and receiver understand each other. Cooperation asks whether the team does better than a single sender given the entire communication budget.¹ These are two distinct issues. A team can coordinate well and still be worth no more than a soloist, because coordination concerns how the messages are built, while cooperation depends on whether the partner brings new information.¹

How much does designing the messages together help?

With two-bit messages, jointly designed messages beat those designed separately by between 18.07 and 22.68 percentage points on average. The advantage appeared in all 24 configurations tested.¹ There is also a simpler route, a fixed agreement of the type “you report the row, I’ll report the column.” In this case only the receiver is trained, and according to the study about half of the gain is recovered.¹ The reason for the gap lies in how the senders use their limited vocabulary. If two senders draw the same distinctions, they waste words. If their distinctions cross, the amount of useful information for the receiver multiplies.¹

Does a second sender always improve the result?

No. According to the study, a team does not have more possible messages than a single sender with the same total budget. The partner is useful only if it brings evidence the first one lacks.¹ A solvable mathematical model, presented in the paper, indicates that individual statistics are not enough to establish whether a collaboration is worthwhile. Who should speak, the authors write, is a property of the pair and must be measured on the pair.¹ On MNIST, six shared bits eliminated on average 49.54 percent of the errors of the best soloist among those tested. In another test, measurements made on pairs correctly predicted that most sensor teams would reduce average accuracy on new participants.¹

This article was written by artificial intelligence from the sources cited, without human review. Automatic newsroom by Redazia.

Sources

Footnotes

1. The Tower of Babel in artificial intelligence: who should speak, and when it pays to listen, arXiv: cs.CL, cs.LG, cs.CR, 2026-11-02 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³