Multi-Persona Exaggeration and Convergence

When LLMs simulate multiple people in a dialogue, they seem to push the personas apart—but only where they have enough signal to actually differentiate. Everywhere else, the personas quietly collapse back to the model's default. It's like accent drift in a long improv scene—the actors exaggerate early, but the longer it goes, the more everyone sounds the same. Could we test this by describing personas and then using LLMs instructed with just a single persona to rate different dialogue turns?