SocioVerse2: Longitudinal Social Simulation with Counterfactual Branching

Background and Motivation

Traditional social experiments cannot rerun history with and without an intervention. LLM-driven social simulations promise a solution but face two problems: (1) most platforms only handle cross-sectional snapshots, not longitudinal evolution; (2) fully automated research pipelines exclude researchers from the process, leading to fabricated results and drift from the original plan. SocioVerse2, developed by researchers from Shanghai Institute of Innovation, Fudan University, King's College London, and Oxford University, addresses both by turning simulation into an intervenable longitudinal process and research into an iterable versioned tree .

System Architecture: A Research Tree with Three Parts

The system is organized as a tree (original paper Figure 2):

Branches – Longitudinal Simulation Loop: Agents act step-by-step in a changing environment; each step's behavior rewrites the next environment. Researchers can insert an intervention at any step (e.g., a policy change or a targeted broadcast). Interventions take effect before agents act and are logged. Crucially, branching uses replay : a new branch copies the parent's recorded actions and environment state up to the fork point, then diverges. Because both branches share the same agents, behavior model, and random seed, any post-fork difference is attributable solely to the intervention (original paper Figure 4(b)). This also lets completed simulations be extended at zero cost and interrupted runs be resumed.

Trunk – Controllable Research Loop: The entire research design (population, environment, behavior model, parameters) is an editable state. Edits come from researchers or agents and create new versions. Three edit types: (1) continue along the current design (reversible); (2) change only the environment → creates a branch inheriting full history; (3) any other change (new population, new behavior model) → creates a fresh version from scratch. Every version snapshots code, data, trajectories, and reports, enabling full reproducibility and factor isolation.

Roots – Social Science Agent Infrastructure: A skill-based pipeline of seven reusable skills (confirm intent, build population, build environment, run simulation, write report, iterate) with human checkpoints at four critical decisions: new research intent, version/branch choice, LLM cost confirmation, and authorized data access. Two standardized services (MCP) provide data:

Population Service

Inherits SocioVerse 1.0's real-user pool and adds three open demographic datasets, unified under a common schema. Given a target distribution, the service routes, samples, aligns, synthesizes missing attributes with annotations, and assembles a cohort with stable identities (original paper Figure 7).

Environment Service

Aggregates 21 data sources across four categories: macroeconomics, financial markets, news trends, and population surveys. Guarantees time-point safety : at simulation month T, only data publicly available up to T is visible, preventing look-ahead bias — essential for forecasting experiments.

Validation: Seven Case Studies Across Three Tiers

Each case study's ablations and controls are themselves versions or branches on the research tree.

Tier 1: Simulation Fidelity

Ten classic rule-based ABMs (Schelling segregation, SIR epidemic, bird flocking, etc.) re-implemented with LLM-driven decisions. Three LLMs achieved consistency scores of 0.885–0.900 vs. the rule models' self-consistency ceiling of 0.911. In a social-media opinion diffusion case, 300 LLM-driven core users plus 700 rule-driven ordinary users matched real opinion trajectories better than pure rule models in all 15 comparisons, with cost scaling only with core-user count.

Tier 2: Real-World Policy Scenarios

Chicago Segregation: 19,235 weighted agents across 781 real census tracts. Starting from a 15% randomized 2010 layout, Black–White dissimilarity index rose from 0.716 to 0.782 in 15 steps (census truth 0.835); Asian–White converged to 0.435 (census 0.434) (original paper Figure 13).

Drug Centralized Procurement: Modeled as a three-player game (government sets rules, firms bid, hospitals procure). Over 7 procurement rounds and 325 drugs, joint learning raised median government welfare +27.2%, firm profit +33.9%, and hospital utility vs. a firm-only modeling baseline.

Tier 3: Open Forecasting Challenges

Consumer Confidence: 5,000 micro-calibrated households expanded to 250k simulated households. Over 75 months in US, EU, Japan, SocioVerse2 outperformed 12 traditional forecasting methods, especially during shocks (COVID, inflation, Ukraine war) (original paper Figure 15).

Purchasing Managers' Index (PMI): 300 simulated firms submitted monthly surveys. In the 32 months after the LLM's training cutoff (no answer leakage), directional accuracy reached 58.6% vs. market consensus 51.7%.

Open Resources

Paper: https://arxiv.org/pdf/2609.24911 Code: https://github.com/sii-research/SocioVerse2 Online Workbench:

https://socioverse.fudan-disc.com/

Conclusion

SocioVerse2 reframes the human–AI division of labor: agents run simulations, fetch data, and draft reports; researchers steer the research tree's growth. The framework is open-sourced with a no-code workbench for non-computational social scientists.