LLM agents spontaneously form authority hierarchies in long-term simulations
Researchers watch AI agents evolve labor specialization and social stratification...
Deep Dive
Zhiyuan Ji and five co-authors propose CAREB-MAS, a multi-agent framework grounding LLMs in Affect Control Theory and Social Identity Theory. Across long-horizon simulations (37-page paper, accepted at ACL 2026), agents spontaneously reproduce five core phenomena of Fei Xiaotong's Differential Order—stable labor specialization, guanxi-based economic ethics, relational decay of cooperation, emergent relational authority, and clan-based center-periphery stratification—with only minimal interaction protocols specified by the macro environment.
Key Points
- CAREB-MAS framework uses LLM agents with emotion-ethics-belief chains and evolving egocentric identities
- Long-horizon simulations (37-page ACL 2026 paper) spontaneously reproduced 5 core Differential Order phenomena
- Patterns shifted from kin-centered integration to functional interdependence depending on production structure
Why It Matters
LLM simulations can now model long-term social structures, offering a new lens for studying human societies.