AI Research Agents Narrow Scientific Exploration

Researchers tested four AI research-agent frameworks and six LLMs on 37,802 ideas drawn from shared AI/ML literature.

Compared with human papers and follow-on work, AI-generated ideas were more clustered and stayed closer to the source literature.

The ideas were associated with lower-cited papers and tended to differ mainly through method recombination, not new research questions.

The results suggest current AI research agents may be better at accelerating incremental work than driving broad scientific discovery.