Researchers have introduced S$^2$-MAS, a multi-agent simulation framework that scales to 500 million synthetic household personas derived from real survey data. The system is designed to model economic responses at society scale, with each agent making decisions that reflect its household circumstances and the effects of earlier actions.
A central design choice is the Genese engine, which separates decision-making from state updates, while reusable financial scenarios and application ledgers track consumption, cash, debt, and payments. The compact decision model, FinL-mlx, was trained on 2.93 million examples and achieved 74.55% accuracy and a 38.37% question-macro F1 in consumer financial-choice scenarios, outperforming both Jev 1.13 and GLM-5.3-Flash.
The authors evaluated eight financial task families and simulated 12 million distinct household agents in paired purchase-funding conditions. Jointly raising purchase goals and approved credit limits increased consumption, but responses differed across age and income groups. After standardizing for age and income, differences tied to education, employment, and household size persisted, reflecting transitions from deferral to purchase rather than simple ability-to-pay effects.
Asset and debt groups also differed in cash drawdown, repayment, and closing debt. A 24-month panel that varied purchase goals under a common credit product showed repayment and arrears updating subsequent offers. The framework is intended to link heterogeneous household decisions to evolving consumption, cash, and repayment paths, supporting economic experiments across household groups, population scales, and time horizons.