A potential approach to enhance idea generation.

[GuoQi2000]

Nice work! I'm impressed by the automatically generated paper!

I’m working on a project called “AutoSurvey” which aims to generate comprehensive literature surveys. I think it complements what you’re doing quite nicely—while scientific discovery focuses on scientific discovery, we’re tackling the summarization and review of existing research.

Based on my experience in AutoSurvey , I would like to suggest a possible approach to help idea generation. According to the process described in your paper, AI-scientist first brainstorms ideas based on its own knowledge and then refines the ideas iteratively. However, a general research paradigm typically involves extensive reading of existing literature to discover new ideas. It seems that the latter approach (such as using RAG techniques to hierarchically reading hundreds of papers) could better utilize LLMs' capability to process long texts.

Have you ever considered this approach?

[conglu1997]

Thank you! That work is very interesting, and we're adding it to our related work section. I think that would be a great idea to incorporate an extensive literature survey, and I wonder how the ideas generated would compare. Please do feel free to propose a change to our generate_ideas file! :)

[erkinalp]

keep in mind the approach suggested by guoqi2000 (first searching the literature and then bruteforce generating novel ideas) is computationally more expensive than the current approach (first discussing internally then looking up the literature to check whether the initial idea is novel)