Idea Generation Agent (Optimization)

[BradKML]

Borrowed from this issue of AI Scientist SakanaAI/AI-Scientist#19
Paper idea generation tools (see also https://github.com/Paureel/LLM-SCI-GEN for more papers)

• https://github.com/EagleW/PaperRobot
• https://github.com/cheerss/SciPIP
• https://github.com/artificial-scientist-lab/SciMuse

Something worth considering is how they parse papers differently than just dumping them in as a pure article.

[SmartManoj]

Should it be another agent like CodeActAgent? or Should CodeActAgent use these tools?

[BradKML]

For some reason having another agents would be best when it comes to creative thinking, like internal dialogues but also have stronger preference for RAG work (e.g. having a list of papers or articles provided).

[SmartManoj]

You could change the system prompt for creative thinking.

Why should RAG be used here? Shouldn't the LLM read the Full paper?

Clean long PDFs for LLM inference

[BradKML]

Thinking RAG is needed if the library has more than a full list of more than a dozen writers, each having like more than a dozen books and papers, so imagine >100 papers blowing up the context window. In the case of pure idea generation rather than topic-centric generation, most people still have >100 papers but spread across a larger set of writers.

[SmartManoj]

Still, should the agent process all the PDFs fully or only a subset of it?

[BradKML]

All of the PDFs assuming the selection is bespoke, but a subset of information (or a summary of information) for each of the PDFs. RAG and/or CoD would be needed.