'When Is Routing Meaningful?' Fewer Than 10 Agents Recover Most Diversity in Model Societies

arXiv·medium signal

The paper introduces two diagnostics — Hierarchic Social Entropy (HSE) for behavioral differentiation among models and a perturbation-based robustness metric — to test whether multi-model routing is actually meaningful. It finds HSE has strong diminishing returns, with a curated subset under ten agents recovering most available diversity, and that KNN routers gain accuracy with specialist models but collapse under perturbation while prompted routing stays stable. A caution that high task accuracy does not guarantee robust routing.