SocioVerse2 released: an open-source agent infrastructure for long-horizon social simulation

Suppose you want to know whether a media campaign can reverse public opinion that is already fermenting. In reality this experiment is nearly impossible to do: you cannot make the same group of people live through the same stretch of experience again, once with the campaign deployed and once with nothing done, and then compare the two outcomes. Large-model-driven social simulation appears to offer a way out, but the research team points out that two problems arise when actually using it for research.

The first problem is time scale: most platforms excel at the "cross-section"—giving a group of agents a questionnaire and seeing whether the distribution of answers resembles a real population; but social phenomena evolve slowly over time, and what researchers want is to insert an intervention at some moment in the same stretch of history, then compare it with the history that had no intervention. The second problem is the human's position: more and more systems hand the entire research process to agents to complete end-to-end, with researchers only taking a look at the report at the very end, making it hard for people to intervene midway when the simulation goes off track or the setup has problems.

These two judgments come from the research team's own paper and official release, and are the source party's claims rather than independently verified conclusions. They constitute the design premises of SocioVerse2: what social science needs is not just a more automated social simulator, but a way of doing research in which humans and agents evolve together.

SocioVerse2 uses a "research tree" to organize the entire process. The branches are the longitudinal simulation loop: the same group of agents acts step by step in a constantly changing environment, and the behavior at each step in turn rewrites the environment for the next step; researchers can insert an intervention at any step, causing the simulation to split off a counterfactual branch. The trunk is the controllable research loop: the entire study—which group of people, what environment, which behavior model, what parameters—is treated as an editable state, and each time a modification is proposed, the research enters a new version. The roots are the social science agent infrastructure: a set of skill-based research pipelines, plus two services that provide real population and real environment data.

This structure follows a classic judgment from Lewin's field theory: a person's behavior is determined jointly by this person and the environment in which they are situated. The system therefore unfolds around the two independently replaceable parts, "population" and "environment," plus a "behavior function" that maps the two into behavior.

It should be noted that the above framework description comes from the paper's own account, and whether its actual operational effects match the design intent is not provided with an independent evaluation by the source.