Do frontier LLMs actually explore the full space of valid answers?

The paper's starting point is that a model can be accurate and still "epistemically narrow": in contexts where alternatives would be useful, it "repeatedly presents only a few canonical routes despite the existence of multiple valid answers." The authors define epistemic diversity as the range of valid answers, explanations, examples, concepts, and reasoning strategies a model makes available to users. Operationalizing it across ten models and two datasets, they find that frontier LLMs "often exhibit epistemic narrowness, even when many valid alternatives exist." In the professions domain the models concentrate on a small set of canonical individuals; in the proofs domain they often return the same proof strategy although accepted alternatives exist.

The reasoning rests on how LLMs are now used: not only to predict but to answer questions, explain, teach, and support inquiry. In those settings what matters is also which answers and reasoning routes the model puts in front of the user. Accuracy cannot register the loss, because each answer a narrow model returns is valid; the narrowing shows up only when you ask how much of the valid space is covered. The paper therefore separates epistemic diversity from group-based diversity, which asks whether demographic, cultural, or political groups are represented or treated fairly. Epistemic diversity concerns "coverage over valid answer spaces or answer classes." The proposed framework asks what makes an answer valid, why multiple valid answers arise, and how diversity should be measured under different interaction protocols. The authors' conclusion is that models "do not merely answer questions, but shape which knowledge becomes salient."

Against the nearest notes, this is a coverage-of-valid-answers reading of a problem the vault already tracks as output homogeneity. Do different AI models actually produce diverse outputs? studies open-ended queries with no single ground truth; this paper's two domains are ones where accepted alternatives exist, so the narrowing is measured against a set of valid answers rather than against human variety. Does preference tuning actually reduce the diversity of model outputs? also gates diversity on quality, but asks how varied the passing outputs are, not how much of the valid space they reach. Why do LLMs generate novel ideas from narrow ranges? shows the same shape in a third setting, where individual quality coexists with a narrow collective range.

The excerpt is silent on much that would sharpen the claim. It does not name the ten models or the two datasets beyond their domains. It does not say how validity was determined, how narrowness was quantified, or how large the effect is. It does not say whether the concentration appears within one model across samples, across models, or both, so it cannot be set directly beside the hivemind finding. It offers no cause, such as training method or scale, and no human baseline. The framework is called preliminary, and only two domains are tested. What follows at that strength is modest: accuracy alone is an incomplete evaluation for knowledge-support uses, and coverage of valid answer spaces is a measurable candidate for the missing dimension.