Signal SocioVerse2: A Longitudinal Dynamic Social Simulation Framework under a Human-AI Co-evolutionary Paradigm

Summary

The paper starts from the premise that generative agents, functioning as silicon samples that unite agent-based modeling with real behavioral data, have reshaped social simulation. It argues that existing platforms mainly verify collective behavior or align simulated populations with real-world snapshots, leaving two needs unmet in a systematic way, namely intervening in what a simulation contains and letting researchers directly control the process that produces it. To address this, the authors present SocioVerse2, which extends the earlier SocioVerse 1.0 into a human-AI co-evolutionary paradigm built around two loops, one that simulates a target population under an evolving environment and forks counterfactual branches through interventions, and another that treats the study itself as an editable state whose versions are updated through controlled editing. Both loops run on a shared infrastructure combining composable skills with researcher checkpoints, a population service drawing on five persona pools, and an environment service that draws on 21 real-world signal sources with point-in-time guarantees. The team validated SocioVerse2 across three case families and seven case studies, ranging from reproducing classic agent-based models to modeling policy processes from real records and forecasting macroeconomic indicators beyond the underlying model's training cutoff. Code, data services and the workbench have all been released as open-source resources.