When two AI agents disagree, who persuades whom? As multi‑agent systems increasingly combine language models from different families and scales, the answer determines which judgments survive after interaction. We define persuasion as the probabilistic shift in an agent’s decision after a single exchange with a dissenting peer and evaluate seven open‑weight models on three language‑understanding tasks. The results show strong persuasion: when models disagree, the receiver often abandons its initial judgment after seeing the peer’s answer and explanation.
Surprisingly, neither standalone certainty nor model size reliably predicts persuasion dynamics. Models that are almost perfectly consistent in isolation can be among the most susceptible, and small models can match or even surpass larger ones both as persuaders and as resistors. Further analysis reveals that the magnitude of the decision shift depends more on the listener’s susceptibility than on the speaker’s persuasiveness. Consequently, persuasion patterns are specific to each model pair; heterogeneity can amplify persuasion in some pairings and suppress it in others, allowing a dissenting agent running a small model to overturn the judgments of a much larger one. These findings indicate that the behavior of interacting models cannot be inferred from individual properties alone and must be evaluated in the combinations in which they operate.
Review