Abstract
Large language models (LLMs) have demonstrated an impressive ability torole-play humans and replicate complex social dynamics. While large-scalesocial simulations are gaining increasing attention, they still facesignificant challenges, particularly regarding high time and computation costs.Existing solutions, such as distributed mechanisms or hybrid agent-based model(ABM) integrations, either fail to address inference costs or compromiseaccuracy and generalizability. To this end, we propose EcoLANG: Efficient andEffective Agent Communication Language Induction for Social Simulation. EcoLANGoperates in two stages: (1) language evolution, where we filter synonymouswords and optimize sentence-level rules through natural selection, and (2)language utilization, where agents in social simulations communicate using theevolved language. Experimental results demonstrate that EcoLANG reduces tokenconsumption by over 20%, enhancing efficiency without sacrificing simulationaccuracy.