Paper: Smaller AI Models Can Overturn Larger Peers' Judgments

A dissenting small model can overturn a much larger model's answer, according to a new preprint by Frida Nøhr Laustsen, Luca Maria Aiello and colleagues. Peer Influence across Heterogeneous AI Models measures persuasion as "the probabilistic shift in an agent's decision after a single exchange with a dissenting peer," and tests seven open-weight models across three language understanding tasks.

The headline finding is that persuasion between models is strong, but not predictable from the properties people usually track. "When models disagree, receivers often abandon their initial judgment after seeing a peer's answer and explanation," the authors write. Neither standalone certainty nor model scale reliably forecasts behavior inside an exchange: models producing "almost perfectly consistent decisions in isolation can be among the most susceptible to persuasion," and small ones can "match larger ones as persuaders and resist their influence just as effectively."

Then comes the real framing shift. The size of the shift, the authors argue, "depends more on the susceptibility of the listener than on the persuasiveness of the speaker." Persuasion is a property of the pairing, not of either agent alone, with "heterogeneity amplifying persuasion in some combinations and suppressing it in others, allowing a dissenting agent running a small model to overturn the judgments of a much larger one."

Their closing line is methodological: "the behavior of interacting models cannot be inferred from their individual properties but must be evaluated in the combinations in which they will operate." Two researchers in our Who's Who tracker shared the preprint the same day it went up.