Large language models are now being placed in multi-agent environments where they often produce unanimous agreement. Sometimes this consensus is the goal, such as when coordination is required; other times it arises on its own from the agents' interactions.
A new preprint on arXiv presents a simulation study of this phenomenon. The researchers model LLM agents with competing incentives to see what pushes them toward agreement. The work focuses on the drivers of consensus formation under such pressure, rather than assuming that agreement simply emerges from shared training data.
The study is notable because unanimous agreement among LLM agents can be both useful and risky. Simulating competing incentives may help clarify when consensus reflects genuine coordination and when it might mask disagreement or blind spots.