Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems

Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems. By Yiqiao Jin and ten others, from the Georgia Institute of Technology, the University of California, Los Angeles, and two other institutions. Published on the first of October. Its code repository is worked out: about one hundred and twenty source files, with a readme and a licence.

Scientific progress emerges from a longitudinal ecosystem. Researchers, institutions, funding agencies, collaboration networks and the scientific literature all co-evolve. As AI becomes increasingly involved throughout the research cycle, understanding those interconnected and evolving processes becomes increasingly important. The difficulty is in the word interconnected. Review, publication, citation, collaboration and funding all feed into one another, and because the ecosystem is longitudinal, their effects unfold across years rather than in a single round.

SciUtopia is a persistent, closed-loop simulation framework built from LLM agents, for studying academic research ecosystems as a whole. It models interconnected processes: choosing a research direction, collaboration, submission, peer review, resubmission, citation, funding, and researcher attrition. It maintains evolving states across simulated years, so what happens in one year carries into the next.

Its institutional mechanisms and information channels are configurable. That is what turns the simulation into a controlled testbed, one where matched counterfactual experiments and targeted interventions can be run. The scale is large. Across sixty-one simulation worlds, SciUtopia simulates over forty thousand researchers from eight thousand institutions. Those worlds produce around four hundred thousand publication decisions, and one point two million LLM-generated peer reviews.

Three findings come out of these longitudinal simulations. Rejection-driven resubmission substantially amplifies the burden on reviewers, beyond what population growth alone would produce. Cautious exploration balances citation impact against career success and long-term topic diversity. And resource inequality can emerge even without detectable cumulative advantage from narrowly winning early funding. Each of these is a statement about consequences that build up over time: a reviewing load that compounds through resubmission, an exploration strategy judged over whole careers, and an inequality that appears even though the advantage of an early, narrow funding win could not be detected.

What is on offer is a place to ask counterfactual questions about science itself. Change an institutional mechanism or an information channel, keep the rest of a world matched, and watch the consequences play out across simulated years of careers, reviews and citations.