SocioVerse-ABM v0.2.0

SocioVerse-ABM v0.2.0 is the first public release of the benchmark that replaces rule-based agents with LLM agents in classic agent-based models and checks whether the collective behaviour is reproduced.

What is included

• 12 tasks with a shared kernel (socioverse_abm): flow (NaSch, Boids, Social Force), market (Sugarscape, Lux-Marchesi, Minority Game, Axelrod), organization (Schelling, Civil Violence, Chicago segregation) and diffusion (SIR, Hegselmann-Krause).
• For each task: the rule-based behaviour function, the LLM behaviour function, a hybrid mode, an evaluation script, and the original implementation under legacy/.
• The Chicago segregation model with real census-tract data (see tasks/organization/chicago_segregation/legacy/DATA_LICENSE.md: the HOLC fields are CC BY-NC; the code is Apache-2.0).
• sv-abm list / sv-abm run <task> --mode rule|llm|hybrid.

Use it with SocioVerse2: clone this repository next to SocioVerse2 (folders named SocioVerse2 and SocioVerse-ABM) to run the 11 abm_* studies and chicago_schelling there.

Technical report: https://arxiv.org/abs/2609.24911 · Homepage: https://socioverse.fudan-disc.com/

---

SocioVerse-ABM v0.2.0 是该基准的首个公开版本：在经典 ABM 中用 LLM 智能体替换规则智能体，检验群体行为能否复现。包含 12 个任务、共享内核、每个任务的规则/LLM/混合行为函数与评测脚本、legacy/ 中的原始实现，以及带真实普查数据的芝加哥隔离模型（数据许可见 tasks/organization/chicago_segregation/legacy/DATA_LICENSE.md）。与 SocioVerse2 并排克隆即可在 SocioVerse2 中运行 abm_* 与 chicago_schelling 研究。