Understanding how individual researcher behaviors aggregate into collective scientific out- comes remains a central challenge in the science of science. We introduce Scientific Liquidity Agents (SLA), a novel multi-agent framework that leverages Large Language Models (LLMs) to simulate researcher behavior in a scientific knowledge market environment. Drawing inspiration from financial market simulation frameworks, SLA applies market concepts—knowledge liq- uidity, friction, and epistemic arbitrage—to model the scientific ecosystem. At the micro-level, we employ the Belief-Desire-Intention (BDI) cognitive framework to model heterogeneous re- searcher agents with diverse skills, behavioral tendencies, and incentive functions. At the macro- level, we simulate a dynamic knowledge graph and collaboration network that enables informa- tion exchange and idea propagation.Our key innovation is the introduction of Epistemic Arbitrageur Agents—specialized re- searchers who reduce system-wide friction by bridging disciplinary boundaries and facilitating cross-domain knowledge transfer. We validate the framework through a simulation of 10 LLM- driven agents over 100 steps, producing 280 publications with a 78% success rate. Key findings include: (1) emergent role differentiation (Arbitrageurs span 3.0 domains vs 1.0 for Special- ists); (2) productivity clustering (Lag-1 autocorrelation: +0.155); and (3) realistic publication tier distributions. The framework successfully reproduces stylized facts observed in real scientific ecosystems, including volatility clustering and knowledge concentration dynamics.
Koutian WU (Mon,) studied this question.