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

Recent advancements in large language models (LLMs) have sparked optimism about their potential to accelerate scientific discovery, with a growing number of works proposing research agents that autonomously generate and validate new ideas. Despite this, no evaluations have shown that LLM systems can take the very first step of producing novel, expert-level ideas, let alone perform the entire research process. We address this by establishing an experimental design that evaluates research idea generation while controlling for confounders and performs the first head-to-head comparison between expert NLP researchers and an LLM ideation agent. By recruiting over 100 NLP researchers to write novel ideas and blind reviews of both LLM and human ideas, we obtain the first statistically significant conclusion on current LLM capabilities for research ideation: we find LLM-generated ideas are judged as more novel (p < 0.05) than human expert ideas while being judged slightly weaker on feasibility. Studying our agent baselines closely, we identify open problems in building and evaluating research agents, including failures of LLM self-evaluation and their lack of diversity in generation.

Introduction. The rapid improvement of LLMs, especially in capabilities like knowledge and reasoning, has enabled many new applications in scientific tasks, such as solving challenging mathematical problems (Trinh et al., 2024), assisting scientists in writing proofs (Collins et al., 2024), retrieving related works (Ajith et al., 2024, Press et al., 2024), generating code to solve analytical or computational tasks (Huang et al., 2024, Tian et al., 2024), and discovering patterns in large text corpora (Lam et al., 2024, Zhong et al., 2023). While these are useful applications that can potentially increase the productivity of researchers, it remains an open question whether LLMs can take on the more creative and challenging parts of the research process. We focus on this problem of measuring the research ideation capabilities of LLMs and ask: are current LLMs capable of generating novel ideas that are comparable to expert humans?

Discussion / Conclusion. To summarize, we compared research ideas generated by our AI agent with ideas written by expert researchers, and observed the robust finding that expert reviewers rate AI ideas as statistically more novel than expert ideas. In this section, we discuss some high-level questions readers might have and suggest some ways to address them. Question 1: Do these collected expert ideas represent their best ideas? One might argue that these ideas submitted by our idea-writing participants might not represent their best ideas as we discussed in Section 6.1, since most of them came up with the idea on the spot within a short period. In order to address this concern, we have designed an experiment where we will compare AI ideas with papers accepted at top-tier AI conferences. To avoid any possible contamination, we target the upcoming EMNLP 2024 conference, which will release the accepted papers in October 2024. We have generated AI ideas with our agent on 23 topics from the EMNLP Call For Papers page in July 2024 and cached them. We pre-registered our analysis plan which also includes the link to the cached ideas.