This paper introduces a large-scale framework for comparing LLM-generated research ideas with human research taste. For each high-quality human paper, the authors reconstruct a small set of related prior works and prompt LLMs to propose new ideas from their titles and summaries. A two-axis taxonomy profiles opportunity patterns and research paradigms. Across multiple LLMs, generated ideas consistently overrepresent bridge-like opportunities and synthesis methods, while human reference papers show broader distributions of gap framing and contribution construction. The results indicate that capable LLMs can produce reasonable ideas, but their ideation space remains narrower and systematically shifted relative to human researchers.
No heat snapshots are available in the last 24 hours.