An arXiv study reports that the structure of a network did little to predict where simulated groups of language models would end up agreeing. Researchers measured 73,080 controlled exchanges across eight instruction-tuned models, then used the results in simulations of 10,000 agents.
The models sometimes shifted opinions beyond the range of views they had seen, and some groups converged on objectively false claims, the researchers say. In their simulations, each model’s internal bias and how messages were framed mattered more than a network measure called algebraic connectivity.
Why it matters
Teams using multiple language models to discuss claims should not assume that changing how the models are connected will prevent shared errors; the study suggests model biases and message framing may still drive agreement.
