[5L1-OS-30] OS-30

10:00 AM - 10:15 AM JST(1:00 AM - 1:15 AM UTC)

[5L1-OS-30-05] Metadata Analysis and Perspectives on Evolving AI Scientists

〇Yatima Kagurazaka¹ (1. The University of Tokyo)

Keywords:

AI Scientists,Self-Evolving Agents,LLM,Recursive Self-Improvement,AI for Science

Large Language Models (LLMs) continue to develop rapidly, with consistent trends observed across multiple performance metrics. This advancement enhances the capabilities of AI scientists (end-to-end automation of science using LLM agents and related technologies) and self-evolving agents, potentially triggering critical challenges through large-scale automation of academic research and recursive self-improvement.
To assess the current state of these important fields from a novel perspective, this study conducted quantitative analysis using paper metadata. We confirmed an increasing trend in both the number of publications and citation counts. Using a task-versus-domain matrix, we identified underexplored areas including hypothesis generation and literature review applications, as well as general applications in mathematics and physics. Additionally, we extracted and analyzed future work sections from the collected papers. This analysis revealed that in hypothesis generation research, advanced reasoning capabilities and architectural innovations tend to be prioritized over validation and rigor. Related materials and updates for this study will be made available at http://yati.ma/scimeta-quant/.