Fudan-led team open-sources SocioVerse2, an intervention-ready LLM social simulation platform
机器之心 · wechat · 2026-10-05
A team from Shanghai Innovation Institute, Fudan, KCL and Oxford released SocioVerse2, upgrading LLM social simulation from one-shot cross-sectional surveys to longitudinal experiments where researchers can insert interventions and compare counterfactual branches. Code, paper, and an online workbench are open.
- A "research tree" architecture: simulation branches fork via replay — history before the fork is replayed exactly, so post-fork differences are attributable solely to the intervention; researchers hold edit rights over the research state, with every change versioned and reproducible.
- Infrastructure layer ships seven reusable skills plus population and environment MCP services (21 signal sources) with point-in-time safety to prevent lookahead.
- Validation across seven cases: LLM agents match rule-based ABM consistency (0.885–0.900); 19,235 agents reproduce Chicago's segregation index (0.716→0.782 in 15 steps vs census 0.835); drug-procurement game lifts welfare and profit 27.2%/33.9% over the strongest baseline; consumer-confidence forecasts beat 12 traditional methods across 75 months, and PMI direction accuracy hits 58.6% vs consensus 51.7%.
More from AGI Musings
- YOLO-mode LLMs create an absurd capability gap against IT-deficient counterparties — curious_vii · 2026-10-06
- AI-Discovered Algorithm Delivers First True Subquadratic 3SUM and Subcubic APSP — FrnkNlsn · 2026-10-06
- Should AI Triage Mental Health Crisis Messages? The Case for Human Review — yi111 · 2026-10-06
- Neuroscientist argues experimental neuroscience can't crack intelligence, but aids disease therapies — aran_nayebi · 2026-10-06
- Bryan Johnson: the global era of Don't Die begins in 2030 — Zachly · 2026-10-06
- Who believes in fast foom but near-zero doom probability, with deep technical arguments? — burny_tech · 2026-10-06