Multi-Agent Social Simulation: An Experimental Framework for Language-Native Social Experiments

胡晓李, yang shen, Keke You

Keywords: Multi-Agent Social Simulation; Research Method; Social Experiment; Social Simulation; Intersubjectivity

TL;DR: We introduce MASS, a novel language-native multi-agent methodology and validate its reliability and validity using a minimum-wage simulation; the paper therefore establishes MASS as a reproducible platform for future ex-ante policy experiments.

Abstract: We introduce Multi-Agent Social Simulation (MASS), a language-native experimental framework in which large language models (LLMs) act as intersubjective agents. MASS models social processes directly in natural language and executes policy shocks in a controllable, round-based environment while recording both numeric actions and one-sentence intention rationales. We specify a minimal, sufficient protocol (time granularity, exogenous rules, agent autonomy, controlled contrasts, independent replications) and use intention texts to link dialogue chains to causal chains for mechanism audits. As a validity and reliability testbed, we revisit the New Jersey–Pennsylvania minimum-wage case: across 46 weekly rounds and multiple independent replications, MASS reproduces the classic Card–Krueger pattern — post-policy starting wages rise (about $0.65/h), employment effects are statistically indistinguishable from zero, and prices move only mildly; a counterfactual group without the wage floor shows no such pattern. Keeping a single modeling equation (the conditional language-model objective) anchors outputs in the public-language corpus that embodies intersubjectivity. MASS combines experimental control and repeatability with language-based coordination and explanation, offering a low-risk platform for ex-ante policy probing and extreme counterfactuals in the social sciences.

Submission Number: 130