MASim: Multilingual Agent-Based Simulation for Social Science

TL;DR

MASim introduces a multilingual agent-based simulation framework to model cross-lingual social interactions and information diffusion. The MAPS benchmark grounds diverse sociolinguistic agents in realistic demographics and survey data, enabling calibrated analyses of global public opinion and media effects. Through experiments and cultural case studies, the framework demonstrates the importance of multilingual simulation for studying sociocultural dynamics, assimilation, and normative diffusion. The results highlight MASim’s potential as a scalable tool for computational social science while acknowledging limitations related to realism and model biases.

Abstract

Multi-agent role-playing has recently shown promise for studying social behavior with language agents, but existing simulations are mostly monolingual and fail to model cross-lingual interaction, an essential property of real societies. We introduce MASim, the first multilingual agent-based simulation framework that supports multi-turn interaction among generative agents with diverse sociolinguistic profiles. MASim offers two key analyses: (i) global public opinion modeling, by simulating how attitudes toward open-domain hypotheses evolve across languages and cultures, and (ii) media influence and information diffusion, via autonomous news agents that dynamically generate content and shape user behavior. To instantiate simulations, we construct the MAPS benchmark, which combines survey questions and demographic personas drawn from global population distributions. Experiments on calibration, sensitivity, consistency, and cultural case studies show that MASim reproduces sociocultural phenomena and highlights the importance of multilingual simulation for scalable, controlled computational social science.

Figures (18)

• Figure 1: The MAPS dataset. (a) User personas derived from WVS, defined by eight socioeconomic attributes plus country and native language; examples from Japan and Canada. (b) Three survey questions with answer options and selected country-level response distributions, illustrated for Argentina, Brazil, and Zimbabwe.
• Figure 2: The MASim framework. Starting from the survey question, in warm-up round $0$ news organization agents create self-introductions to augment their profiles and write posts stating their editorial stances, which are then fed into the recommendation system. User agents likewise create self-introductions, write posts to express their perspectives as input to the recommendation system, and participate in a multiple-choice vote. In rounds $t \geq 1$, news organization and user agents first read recommended posts, then write new posts for recommendation based on their memories and the content they read, and finally user agents vote to update the attitude distribution.
• Figure 3: Experimental results. (a) Lower RMSE scores correspond to better real-world calibration. (b) Smaller RMSE variance indicates more stable simulation outcomes. (c) Larger distribution shift captures more pronounced stance changes induced by media effects. (d) Higher consistency scores represent better quality agent actions.
• Figure 4: Case study. Subfigure (a) shows cultural assimilation attitude changes for Cases 1 and 2, and normative diffusion attitude shifts for Case 3. Semi-transparent dashed lines correspond to simulations without cross-cultural communication. Subfigure (b) depicts, for each country, the maximum share of user recommendations coming from any single foreign country (i.e., Dominant Foreign Exposure Ratio). Markers denote this largest foreign country by the initial letter of its name, while a value of zero indicates that no foreign posts are recommended in that round.
• Figure 5: Real-world calibration by model and country.