Groups of LLM agents agree on incorrect claims in each network shape

Researchers tested how eight instruction-tuned language models change their view after they read the view of others. They used 73,080 controlled exchanges, then simulations of 10,000 agents. The shape of the network changed the last result by a small amount. Some models, with a maximum of 7 billion parameters (model values), agreed with confidence on claims that are objectively incorrect.

How it works

The team read a number for belief from the token probability values of each model. They fitted how each model updates that number. Classic theory assumes that an agent moves between the views it sees.

What the test found

• The models moved to views away from the range they saw.
• Some moved away from the view of others.
• The update bias of the model and the words of messages set the result.

What is not known

The large runs are simulations that use fitted models, not live systems. The authors call the risk collective delusion. The abstract tests no procedure to stop it.

Insight: The model and the message set the result, not the network

Earlier tests found that six open models got consensus less frequently than humans, and that thousands of agents moved into opposite camps. This paper adds that network shape does not set the result. This reads as a move of control: who picks the model and the message words sets the outcome.