Do AI research agents explore as broadly as human researchers?

Across five agent frameworks (a zero-shot baseline, AIScientist, ResearchAgent, AgentLaboratory and Co-Scientist) and five language models, the authors generate 219,655 valid ideas across 155 research areas in 12 fields, then set them against human papers. Within an area, AI ideas are more similar to one another than human-authored papers are: average breadth is 0.554 against 0.599, which the source reports as 7.5% lower exploration breadth. AI ideas also stay closer to their seed literature than the follow-on human papers that cite it, with exploration distance averaging 0.322 against 0.410 across years, a difference the source reports as statistically significant in every field. The abstract adds that the ideas sit in lower-impact regions of the historical landscape and align less with future human research. The overall verdict is that current agents "appear better suited to local elaboration than to broadening scientific exploration."

The paper's account turns on a gap between instruction and output. The prompts push toward novelty: AIScientist aims for "novel" and "high-impact" ideas, Agent Laboratory asks for ideas "very innovative and unlike anything seen before," and Co-Scientist "explicitly rewards novelty at each round." Every framework except the zero-shot baseline can also retrieve further papers from a local Semantic Scholar database, yet the authors read the distance result as showing that the agents "remain largely confined to local exploration." The concentration is not traced to any one component. Ideas from different frameworks sit at a breadth of 0.572 relative to one another, and ideas from different models at 0.570, both below the human same-area baseline, and the pattern holds across all five frameworks, including the multi-agent deliberation and tournament designs.

This extends Why do LLMs generate novel ideas from narrow ranges?, which pairs high per-idea novelty with a narrow collective range. The excerpt measures the narrowness against human papers in the same area and against the literature each idea starts from, across a population of 219,655 ideas, so it tests the collective claim directly rather than inferring it from individual ratings. It is the same compression that Does AI assistance homogenize or preserve creative diversity? describes, here at the scale of a whole field. It also qualifies Can specialized agents write better scientific papers than single models?. That note measures manuscript quality, a different outcome, so the two do not conflict. But the concentration appears under multi-agent deliberation and tournament designs too, so this excerpt gives no sign that multi-agent structure widens the ideas, even where it improves the writing.

The excerpt does not establish everything the abstract claims. It breaks off partway through Section 4.3: the frontier-alignment comparison appears only as comparable keyword counts (11.80 against 11.88), so the finding that AI ideas align less with future research rests on the abstract alone. Neither "impact" nor "valid" is defined in the text, so nothing here shows whether the ideas are sound, or that narrower ideation makes worse science. The breadth and distance scores depend on a similarity measure the excerpt does not specify, and the human breadth baseline is the seed literature itself, sampled one paper per AI idea. The implication, at the strength this evidence allows, is narrower than "AI cannot broaden science": judging AI ideas one at a time, as novelty ratings do, cannot show whether a population of them widens a field, and that population-level measurement is what the authors add.