Why do LLMs plateau on creativity tasks while humans reach further?

This reads the question as "where exactly does LLM creativity stall, and why" — and the surprising answer in the collection is that the plateau usually isn't where people assume. The raw generation of novel ideas is often *not* the weak point. In a controlled study of 100+ NLP researchers, LLM-generated research ideas were rated as statistically *more* novel than human experts' ideas Do language models generate more novel research ideas than experts?, precisely because they lack the disciplinary priors that constrain experts to safe combinations. So if there's a ceiling, it's not a ceiling on combinatorial reach.

The more revealing finding is that creativity is two skills, not one — and LLMs only have the first. Ideation and evaluation are *dissociated*: models combine ideas freely but systematically avoid taking the evaluative stance needed to judge whether an idea is feasible or valid Can LLMs generate more novel ideas than human experts?. This is the mirror image of human experts, whose knowledge constrains novelty but lets them filter. You can see the trade-off play out directly in design work, where LLM solutions score higher on feasibility and usefulness but lower on novelty — and few-shot prompting *narrows* diversity further as it improves quality alignment Why do LLMs excel at feasible design but struggle with novelty?. Push for usefulness and the novelty collapses; push for novelty and nothing prunes the bad ideas.

A second source of the plateau is that current methods only know how to do *conventional* problem-solving. One paper argues creative reasoning actually requires three distinct paradigms — combinational, exploratory, and transformational — and that existing LLM reasoning approaches address none of them, which may directly explain the "diversity collapse" people observe in ideation Can LLMs reason creatively beyond conventional problem-solving?. Humans reach further partly because they fluidly switch between recombining, searching within a space, and breaking the space open; models stay stuck in the first mode.

Underneath all of this is an exploration deficit that looks almost architectural. LLMs commit to answers prematurely because uncertainty signals dominate the early transformer layers while the "empowerment" signals that reward keeping options open only emerge in the middle layers — a temporal mismatch that closes off long-horizon exploration before it can fire Why do large language models explore less effectively than humans?. Even in simple decision tasks, models can't reliably explore without external memory summarization and explicit prompting Why do LLMs struggle with exploration in simple decision tasks?, and reasoning models tend to *wander* rather than search systematically, so success drops off sharply as problems get deeper Why do reasoning LLMs fail at deeper problem solving?. Creativity that "reaches further" is mostly sustained, structured exploration — exactly the thing that breaks.

The deepest framing in the corpus suggests the gap may not be fixable by better sampling at all. One line of work argues LLMs absorb the same shared symbolic system humans do but never develop the *participatory* subjectivity — the reflexive stance of having a position and revising it — that comes from being socialized into the world Do LLMs develop the same kind of mind as humans?. The interesting thing you may not have expected: the human "further reach" might be less about generating more and more about *caring* which ideas are worth keeping — an evaluative, stake-holding act the current architecture sidesteps.