GASim: Graph-Accelerated Framework for Social Simulation

Researchers have introduced GASim, a hybrid multi-agent framework that leverages graph acceleration for extensive social simulations. This new approach tackles the significant latency issues found in earlier hybrid techniques that merge LLM-based agents with numerical agent-based models (ABM). GASim employs Graph-Optimized Memory (GOM) to enhance core LLM-driven agents, shifting from heavy retrieval processes to efficient propagation across a sparse memory graph. For standard agents, it implements Graph Message Passing (GMP), allowing for parallel updates through feature aggregation instead of the traditional sequential ABM execution. Additionally, an Entropy-Driven Grouping (EDG) manages the hybrid partitioning. The research paper can be accessed on arXiv with the ID 2605.07692.