Most AI models secretly think alike, so picking a diverse team means measuring how they actually differ — not just using more of them.

This explores how to pick the smallest set of models that together cover the most behavioral variety, rather than assuming more models means more diversity. The corpus reframes the whole problem: the constraint isn't finding diverse models, it's that models are far less diverse than their number suggests. Studies of 70+ models across 26K open-ended queries find an "Artificial Hivemind" — different LLMs independently converge on nearly identical outputs because they share training data and alignment procedures Do different AI models actually produce diverse outputs?. So a coreset can't be built by counting models; it has to be built by measuring the behavioral distance between them.

The most direct answer comes from work that treats a model pool as a society and measures it with Hierarchic Social Entropy — a metric that scores how behaviorally far apart models actually are. The striking finding: diversity saturates within fewer than ten agents. A curated handful captures most of a large pool's behavioral coverage, and everything past that adds redundancy, not range How quickly does behavioral diversity plateau in model pools?. That's your selection rule in practice: greedily add the model that most increases measured behavioral entropy, and stop when the curve flattens — which happens fast.

But what you measure diversity *over* matters as much as the metric. Measuring raw output variance is misleading, because base models look diverse mainly by spanning incoherent, low-quality space; when you restrict the comparison to outputs that pass a quality bar, the ranking flips Does preference tuning actually reduce the diversity of model outputs?. A useful coreset maximizes *effective* diversity — variety among the answers worth keeping. This is why raw diversity alone isn't the goal: a committee of varied-but-mediocre proposals only pays off when there's a soundness signal (tests, proofs, type checks, an aggregator) to convert the good proposals into selected ones When can weak models match strong model performance?. Diversity is raw material; selection needs an arbiter.

That shifts the objective from "which models are individually best" to "which set is collectively useful." Set-level reinforcement learning makes this explicit — it rewards a batch of traces for complementarity rather than per-member accuracy, and finds that diverse mediocre traces beat redundant strong ones because aggregators need contrast to work with, not confirmation Can diverse mediocre traces outperform redundant expert traces?. The coreset criterion is therefore marginal contribution to the *set*, not standalone rank.

Two more corpus threads sharpen the picks. First, size: for generating unique outputs per sample, small models (~500M params) can beat large ones, because large models concentrate probability mass and collapse variety — so a diverse coreset may deliberately favor smaller, higher-entropy generators Why aren't bigger models better for generating diverse outputs?. Second, provenance: preference tuning changes diversity in opposite directions across domains — compressing it where the domain rewards convergence (code) and expanding it where it rewards distinctiveness (creative writing) Does preference tuning always reduce diversity the same way?. So the right coreset is domain-specific: models that diverge on your task, measured on quality-passing outputs, added until the entropy gain flattens — usually well under ten.