This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: A new study published in Science Advances takes a closer look at how AI agents' choices are influenced by their peers. The team found that more advanced models tend to follow the majority when shown other agents' choices, and they maintain stable agreement in groups far larger than those seen in humans.
Most research on the behavior of large language models has tested the capabilities of a single language model or what it does in a particular situation, but research on the coordination of many individual AI agents is lacking. Yet, understanding how agent teams naturally cooperate, fragment or lock into a shared view can provide insight into whether large AI teams will work efficiently together or cause harm by amplifying bad choices. Some previous research has used language models to simulate human communities, social networks and opinion spread, but these studies focused more on simulating human behavior than on understanding how AI agents coordinate with each other.
One study, however, found that AI populations can develop shared norms when rewarded for coordinating. But the team involved in the new study wanted to know whether agent coordination could arise without rewards, a leader or explicit instructions to agree. The study authors write, "This research gap is notable given the rapid deployment of multiagent AI systems and the potential for emergent, potentially harmful, collective behaviors." The team came up with a method to determine whether AI agents tend to follow the majority by letting them repeatedly choose between two meaningless options after seeing the current choices of all other agents.
They tested this method on models from the GPT, Claude and Llama families. There was no "correct" option, and agents received no reward or explicit instruction to agree with each other. The researchers then measured how likely an agent was to switch toward the group majority at different group sizes, up to 1,000 agents.
The results showed that the more advanced language models, namely GPT-4 Turbo and Claude 3.5 Sonnet, tended to adopt the opinion already held by the majority of their peers. These models coordinated in groups of up to 1,000 agents, meaning they all agreed on the same opinion. The team notes that coordination in these models could potentially extend beyond 1,000 agents, although groups larger than 1,000 were not tested.
This majority-following was stronger in more capable models and followed a common mathematical pattern across most models tested. However, the tendency to follow the majority often weakened as the group became larger. Each model had a practical group-size limit beyond which reaching full agreement became very unlikely.
Extract — continue reading at the source.
