How quickly does behavioral diversity plateau in model pools?

Adapting Hierarchic Social Entropy (HSE) to language-model societies — measuring inter-actor distance by cosine similarity over behavioral vectors — the routing paper finds diversity exhibits strong diminishing returns. On EmbedLLM and RouterBench, a curated subset of fewer than ten agents recovers most of the diversity available in a much larger pool. Beyond that handful, adding models mostly adds redundancy: they occupy behavioral space already covered, so a router gains little by having them to choose from.

This is a practical coreset heuristic, and it inverts a common intuition. The instinct in multi-model systems is that a bigger stable of models is strictly better — more coverage, more specialists. HSE says the behavioral payoff saturates quickly, therefore the design question is not "how many models can I pool?" but "which under-ten models span the behavioral space?" This dovetails with the vault's saturation findings: When does adding more agents actually help systems? finds capability saturating around 45%, and this note locates a parallel saturation on the diversity rather than capability axis.

It also gives teeth to the meaningfulness argument. If routing is vacuous over a redundant society, then curating a small diverse coreset is not just cheaper — it is what makes routing non-vacuous in the first place, because it maximizes behavioral differentiation per model retained. The economic reading aligns with Can small language models handle most agent tasks?: a small, deliberately heterogeneous set beats a large homogeneous one.