Do language model groups mimic human group reasoning patterns?

The paper compares human group chats with matched LLM deliberation traces on Wason-style deductive reasoning. Its core finding is that LLM groups reproduce the aggregate "assembly-bonus asymmetry" from human social psychology: discussion improves the average member more often than it protects the best initial member. The authors argue that this outcome-level match "masks process-level differences." Compared with humans, LLM groups "follow majorities more often, surface less unique information, and converge earlier." Their conclusion is that for human group simulation the mapping between human and LLM group processes is "approximate rather than one-to-one."

The framing rests on a distinction that final accuracy hides. The same negotiated answer can come from an assembly bonus, where interaction corrects initially wrong agents, or from process loss, where interaction pulls initially correct agents away from the right answer. On the paper's account, initial-answer diversity is what explains the effect of model heterogeneity, and it increases movement in both the corrective and the destructive direction. Diversity therefore raises the stakes of the group process instead of guaranteeing a gain. The excerpt also reports that correct minority signals succeed mainly when re-expressed early, and that unique information is realized at a lower rate even though it was available in the group's collective initial state.

This sits alongside the existing multi-agent notes as a human-referenced baseline. When does debate actually improve reasoning accuracy? says debate lifts accuracy where answers can be checked. That is compatible with an average-member gain, but this paper shows that a task with a correct answer still produces destructive movement for initially correct members. Why don't LLM agents naturally explore each other in teams? describes premature commitment and myopic peer interaction. The lock-in reported here is the same family of problem, measured against human groups. Do self-organizing agent teams outperform rigid hierarchies? reports large protocol effects on outcomes. The interventions here are drawn from human group-decision research and give only "modest improvements" without removing "the coordination bottleneck." These are different interventions on different tasks, so the two results are not a direct conflict.

The excerpt does not establish sample sizes, which models were compared beyond a passing mention of GPT-5, effect sizes, what the interventions were, or how the analogical, abductive, and analytical tasks came out. The abstract says the study tests whether the process signatures generalize, and the excerpt does not give that result. It also notes that GPT-5 can reach near-ceiling solo performance on tasks where human solo accuracy is well below perfect (Appx. A). So the human comparison holds only in a "matched task regime." What follows at that strength is a caution for anyone using LLM groups as stand-ins for human groups. Matching outcome-level patterns is not evidence of matching deliberative mechanisms, and the process traces are where the two diverge.