arXiv: AI Research Agents Narrow Rather Than Broaden Scientific Exploration — 37,802 Generated Ideas Cluster Around Existing Work

arXiv·medium signal

Researchers generated 37,802 scientific ideas using four AI research-agent frameworks and six LLMs from shared seed literature across AI/ML research areas. Despite frameworks explicitly encouraging novel and high-impact ideas, the generated ideas tend to narrow scientific exploration, concentrating around existing work rather than exploring new directions. Key implication for builders: AI agent ideation needs deliberate diversity injection to avoid convergent thinking.