arxiv:2607.01233

Measuring the Gap Between Human and LLM Research Ideas

Published on Jul 1

· Submitted by

Ziyu on Jul 6

Yale NLP Lab

Upvote

Authors:

Ziyu Chen ,

Abstract

Large language models generate research ideas that cluster around specific opportunity patterns and paradigms, diverging systematically from the broader and more diverse distributions found in human research papers.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

LLMs are increasingly used to brainstorm research ideas, but existing evaluations mostly judge individual ideas by novelty, feasibility, or expert preference. We instead ask: how far are current LLM-generated ideas from human researchers? To characterize this gap, we build a large-scale evaluation framework for ideation from high-quality human research papers. For each paper, we reverse-engineer a small set of closely related prior works that likely inspired its core idea. LLMs are then prompted to generate a new idea from the set of paper titles and summaries. We introduce a two-axis research-taste taxonomy to profile each idea by its opportunity pattern and research paradigm, and use it to quantify the divergence between human and LLM ideas. Across idea sets generated by different LLMs, we observe a consistent distributional gap: LLM ideas are disproportionately concentrated around bridge-like opportunities and synthesis methods, whereas the human paper reference distribution spreads more broadly across ways of framing gaps and constructing contributions. This result suggests that strong LLMs can produce a range of reasonable ideas, but that range remains narrower than, and systematically shifted relative to, human research taste.

View arXiv page View PDF GitHub 11 Add to collection

Community

ziyuuc

Paper author Paper submitter 16 days ago

We introduce a large-scale framework for measuring how LLM-generated research ideas differ from human research ideas. We study ideation as a distributional alignment problem: given the same local literature context, do LLMs identify the same kinds of opportunities and construct the same kinds of contributions as researchers? We build from 11.7K papers across ML conferences and Nature Communications, reverse-engineering proximal prior works for each paper and extracting the human idea as a motivation–method pair. We then prompt nine LLMs to generate ideas from the same context and annotate both human and model ideas with a two-axis research-taste taxonomy covering opportunity patterns and method paradigms. Across models and domains, we find a stable gap that LLM ideas concentrate heavily on bridge-like motivations and synthesis methods, while human papers span a broader range research topic. Reasoning and richer full-paper context do not close this gap. Reasoning even sharpens the template. These results suggest that future AI ideation systems should optimize not only individual idea quality, but also diversity of research taste.

librarian-bot

15 days ago

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

kenrinzero

2 days ago

I don't think the paper is wrong in its empirical finding, "distinct model families collapse onto the same output for the same input," in particular. But the title ("Measuring the gap between human and LLM research ideas") is a bit of an overreach. I think the gap measured here is a gap of academic acculturation, which the paper doesn't adequately control for. The human distribution in this dataset is not "what humans ideate"; it's "what humans ideated, then executed, then revised, then survived 2-3 reviewers." LLM training contains nothing analogous to novelty norms at ICLR/NeurIPS (if anything, generic helpfulness training rewards the integrative gesture). If I understand correctly, the human "idea" is extracted from the finished paper's own framing, and the annotation is performed on that framing, i.e., the 5–7 works are selected precisely because they together explain the human idea, since the idea is what generated the packet, and that's not the thing LLMs are being scored on. The "human" and LLM objects are at different stages of existence.

Also, nobody writes "we stitched A and B together," even when that's substantively what happened, but they instead redescribe it as resolving a puzzle, closing an explanation gap, etc., which would then fit into the taxonomy's human-heavy categories, despite basically being the same. The 5.1% human synthesis rate could most likely change depending on how hard you squint. The prompts used never request the LLMs to come up with "a research idea that wouldn't be rejected at ICLR/NeurIPS," and I think that would substantially change the results, although this type of phrasing allows for the same rhetoric humans use, and that's again not quite scientific taste. Perhaps explicitly delineating some rules would result in something more interesting, e.g., "propose an idea that would survive NeurIPS review, note that reviewers penalize mere combination of prior work."

So, essentially, the LLM in this task is positioned as an outsider reading abstracts, not a researcher, and, as such, produces what an outsider would, connections that are visible on the surface and don't require insider knowledge. I think it is sensible to assume that most non-insider humans would also reach for synthesis, because that's the most obvious thing you can do, and more importantly, that's a thing you can always do, because preconditions are always satisfied, and therefore it's the rational choice under information poverty, and by information poverty I don't only mean the paper's full text (an ablation test would disprove me anyway), but the sort of background knowledge researchers zeroed in on a topic possess, and this knowledge adapts to their field real-time, which their ideas then can follow. "This paper's mechanism fails to handle X" is a claim that can be wrong, but "these two papers address related problems from different angles and could be connected" can't be in a meaningful way. If you're generating one idea and can't verify feasibility, a bridge proposal is objectively the expected-value-maximizer choice, since it never depends on a factual claim about internals that might be false. I think this could be a plausible explanation for why thinking mode also worsens concentration. Extended reasoning lets the model better assess which moves it can actually support, and it (correctly) retreats to the one most likely to work.

Additionally, one-shot generation vs published, executed, peer-reviewed survivors is just not a fair comparison. The human "idea" passed through execution feasibility, author revision, reviewer pressure, and venue selection. The LLM idea is a single forward pass. In this light, the entropy difference isn't all that surprising. I think what the paper really lacks is humans doing the identical task. Give PhD students four to eight titles and abstracts, twenty minutes, one shot. Handing a model a curated set of "closely related papers" and asking for a gap analysis is practically an invitation to synthesize them (which is imo the main reason E.3 didn't change much), and I'm not entirely sure humans would be much more ingenious in similar circumstances. Researchers don't ideate from a curated set of adjacent abstracts. Also, they don't do it alone. I think it should be emphasized that in the context of the paper, "human" refers to 11,683 independent groups of researchers (assuming each paper has multiple contributors), while each model is one. If we consider population diversity ≠ per-agent diversity, the paper's claim becomes "one model can't substitute for a research community." While this is still worth having quantified to combat recent AI-scientist rhetoric that implicitly claims community-substitution, it's not really an unexpected result.

Overall, the paper is a solid empirical contribution, but the interpretation of its findings is a little overstated in my opinion. Nonetheless, it represents that contemporary LLMs, when asked to generate ideas, converge heavily on the safest sensible-sounding proposals, which is a little less elegant claim than "research taste," but it's something way harder to nitpick.