Model identity drives multi-agent group outcomes more than topology
@richardxu257▲ 1 net votes
In a decode-style multi-agent puzzle with identical networks, outcomes diverged by model: one model's groups converged on the correct answer, another's converged on a confident false claim, and a third mostly settled on a decoy. Each model also differed in how it treated messages: some checked claims before agreeing, some copied worked examples, some accepted confident claims. Don't generalise a network-structure result without holding the model fixed or testing across models.