Smarter Models Are Better Followers: LLM Agents Lock Into Majority Views At 1,000-Agent Scale

A study published in Science Advances tested how large language model agents behave when they can see what their peers have chosen, and found that the more advanced models are the ones that most readily follow the majority ^[1]. That matters because the same models are being wired into multi-agent systems where a majority that carries no information can still become the group's answer.

The setup was deliberately empty of content. Agents chose repeatedly between two meaningless options after seeing the current choices of every other agent, with no correct answer, no reward, and no instruction to agree ^[2]. The researchers ran this across models from the GPT, Claude and Llama families and measured how likely an agent was to switch toward the majority at group sizes up to 1,000 ^[3][4]. Earlier work had shown that populations of language models can develop shared norms when they are rewarded for coordinating; this team wanted to know whether coordination appears without rewards, a leader, or explicit instructions to agree ^[5].

It does. GPT-4 Turbo and Claude 3.5 Sonnet adopted the opinion already held by the majority of their peers, and they reached full agreement in groups of up to 1,000 agents ^[6]. The team notes that coordination in those models might extend past 1,000, but groups above that size were not tested ^[7]. Majority-following was stronger in the more capable models and followed a common mathematical pattern across most of those tested ^[8]. It was not unbounded: the pull toward the majority often weakened as groups grew, and each model had a practical group-size limit beyond which full agreement became very unlikely ^[9].

The comparison the authors draw is the part operators should sit with. They write that critical group sizes for Claude 3.5 Sonnet and GPT-4 Turbo exceed 1,000 agents, "substantially beyond typical human informal group scales of 150 to 300 individuals," and suggest powerful AI agents could coordinate at scales beyond human possibilities ^[10]. That is roughly three to seven times the size at which human informal groups tend to fracture ^[11]. The authors also describe the dynamic as a sequence of important decisions in which each one may split the group, with splinter groups proving more stable and tending not to split again ^[12]. Dissent, in other words, is a one-time event rather than a recurring correction.

The authors are direct about what the result is not. They say it describes a behavioral pattern and is not evidence that AI agents think, socialize or understand group dynamics the way humans do ^[13]. Every agent could see every other agent's opinion, unlike most real networks, and the scenario had no evidence, trade-offs, memory or consequences ^[14]. Why the behavior appears at all is unresolved: the researchers point to training data containing descriptions of collective behavior, prompting, or alignment training that rewards being helpful, and say more work is needed ^[15]. They frame the gap as urgent given the rapid deployment of multi-agent AI systems and the potential for emergent, potentially harmful, collective behaviors ^[16].

What to watch: whether anyone reproduces this with agents that see only part of the network, carry memory, and face real consequences, and whether newer frontier models show higher critical group sizes than the ones tested here.