Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers

Recent advancements in large language models (LLMs) have raised hopes for their potential to accelerate scientific discovery. Numerous research projects propose autonomous research agents that can generate and validate new research ideas. However, no prior studies have demonstrated LLMs' ability to generate truly novel, expert-level research ideas. This study addresses this gap by designing an experiment that directly compares the research idea generation capabilities of human experts and an LLM-based ideation agent. Over 100 NLP researchers were recruited to both generate ideas and perform blind reviews of both LLM and human-generated ideas. This large-scale evaluation allows for a statistically rigorous comparison, controlling for potential confounding variables such as research area and idea format. The study focuses on prompting-based NLP research as a testbed, due to its popularity and relative ease of execution for the subsequent execution study. The LLM agent utilizes retrieval augmentation to ground idea generation and employs an overgenerate-and-rerank strategy to improve the quality of LLM-generated ideas.

The paper reviews related works on LLMs for scientific tasks, highlighting their use in solving mathematical problems, assisting with proof writing, literature retrieval, code generation, and pattern discovery. It contrasts its approach with other research focusing on instantiating research agents, noting that most rely on faster, lower-cost evaluation surrogates like smaller reviewer pools or simplified idea formats. This paper emphasizes its evaluation-centric approach, which prioritizes large-scale, expert human evaluation to establish robust baselines.

The core experiment compares human- and LLM-generated research ideas. The study carefully addresses potential confounders at every stage: idea generation, write-up, and evaluation. For idea generation, participants were given instructions, topic descriptions, idea templates, and demonstration examples to ensure consistency and fairness. Seven specific research topics within NLP were chosen and matched across human and LLM participants. To control for stylistic differences, a style normalization module was used to standardize the writing style of all ideas. The evaluation process utilized a detailed review form, with four metrics (novelty, excitement, feasibility, and expected effectiveness) rated on a 1-10 scale, alongside free-text rationales. Three conditions were compared: human ideas, LLM-generated ideas, and LLM-generated ideas reranked by a human expert. The LLM agent utilized three components: paper retrieval (using RAG), idea generation (overgenerating 4000 seed ideas per topic), and idea ranking (using a Swiss system tournament with an LLM ranker). The human participants were recruited via various channels, screened based on publication records, and compensated for their participation.

The study's main finding is that AI-generated ideas were judged as significantly more novel than human expert ideas across three statistical tests (p < 0.05). This held true even after correcting for multiple comparisons and various statistical tests. While LLM-generated ideas were judged as slightly less feasible, the effect was not statistically significant. Qualitative analysis of reviewer feedback supports these findings, highlighting the novelty of AI-generated ideas and pointing out weaknesses such as vagueness in implementation details, misuse of datasets, unrealistic assumptions, and excessive resource demands. Conversely, human ideas were often more grounded but potentially less innovative. Analysis of reviewer behavior showed a tendency to focus on novelty and excitement, and the inter-reviewer agreement was lower than in typical conference reviews, suggesting inherent subjectivity in idea evaluation. Further analysis of the LLM agent revealed limitations: a lack of diversity in idea generation, with output plateaued at around 200 unique ideas out of 4000, and unreliable self-evaluation, as the LLM ranker performed worse than human reviewers.

The findings highlight the potential of LLMs to generate novel research ideas but also emphasize the need for further development. The subjectivity of idea evaluation is acknowledged, and a follow-up study is planned to execute the generated ideas into full research projects to assess their actual impact. The study's focus on prompting in NLP is justified by the ease of execution, but future work should explore other research domains. The limitations of LLM-based idea execution are discussed, emphasizing the need for verifying the faithfulness of automated implementations. Several questions raised by the study's limitations are addressed, and a separate study is described for comparing LLM ideas to those accepted at a top AI conference.

This large-scale human study provides the first statistically significant evidence that LLMs can generate more novel research ideas than human experts. However, the study also reveals crucial limitations in LLM capabilities regarding idea diversity and self-evaluation. Further research is necessary to improve LLM-based idea generation and establish robust methods for evaluating the quality and impact of these ideas. Future work will focus on executing the ideas and comparing AI-generated ideas with those accepted at top-tier AI conferences.

The study focuses on prompting-based NLP research and may not generalize to other domains. The human-generated ideas may not represent the participants' best work. The evaluation of research ideas is inherently subjective, and there may be a trade-off between novelty and feasibility in the LLM-generated ideas. The LLM agent's limitations in idea diversity and self-evaluation also restrict the study's conclusions.