In recent years, the automation of research using LLMs has been advancing rapidly. While The AI Scientist can generate papers that meet the acceptance criteria of top conferences in the machine learning field under specific conditions, there are limitations to the innovativeness of the generated research. As a step toward improving quality, this study aims to develop a method that generates scientific hypotheses of equivalent quality with significantly fewer tokens. The proposed method, which generates hypotheses more than ten times more efficiently, was compared with previous research in terms of novelty, importance, clarity, feasibility, and validity of the generated hypotheses. While no clear differences were observed in novelty and feasibility, improvements in performance were recognized in terms of importance, clarity, and validity compared to previous research.
Keywords: Hypothesis Generation, Large Language Models, AI for Science
Abstract:
Submission Number: 27
Loading