In this work we empirically measure belief adoption in large language model (LLM) agents, quantifying the probability that an agent accepts a claim as a function of how many peers endorse it. The measured adoption kernel follows a sigmoid shape, a hallmark of complex contagion. The adoption threshold is modulated by three factors: the plausibility of the claim, the reliability of the source, and the agent’s own disposition. These three dimensions collapse onto a single effective dimension that we interpret as the coherence between the incoming belief and the agent’s prior beliefs. Network‑level experiments reveal that belief spread is faster on clustered graphs than on random graphs, reproducing the characteristic of complex contagion. The system also exhibits a bifurcating cascade window and a self‑sustaining hysteretic consensus, making it much easier to establish consensus than to dismantle it.
Review