Why do LLMs generate ideas the research community already explores?

Scientific discovery is constrained not only by what is true but by what is cognitively available to the researchers currently in a field. Many directions are coherent in light of the literature yet unlikely to be proposed, because no existing community occupies the right combination of concepts, methods, and intuitions. The paper's key move is to distinguish two quantities that are usually conflated: whether an idea is scientifically plausible, and whether the current community is likely to think of it.

Modern LLMs are strong at the first and structurally biased on the second. Trained on the literature and prompted through language, they inherit the distributional shape of that literature — they recombine high-density regions, searching where the community has already placed conceptual mass. The result is ideation that is fluent and useful but not complementary to human researchers. The method targets the complementary region — the "alien space" — by decomposing papers into idea atoms, then ranking atom combinations that maximize a coherence model while minimizing an availability model (whether any author community is positioned to produce them).

This is the sharpest available framing of why LLM ideation underwhelms, and it complicates Why do LLMs generate novel ideas from narrow ranges?: diversity collapse is not just homogeneous sampling, it is sampling from the community prior. The honest caveat the paper keeps is that alien ideas are not guaranteed breakthroughs — most research ideas fail. But it relocates the value of AI ideation from "generate plausible ideas" (which the community can already do) to "generate plausible ideas the community structurally cannot reach."