Emergent Relational Order in LLM Agent Societies: From Collective Affect to Authority Stratification

Emergent Relational Order in LLM Agent Societies: From Collective Affect to Authority Stratification

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2606.23764 Date read: 2026-06-24 Connected to: none Escalation: store-only Escalation rationale:

What this is

A multi-agent simulation framework (CAREB-MAS) that grounds LLM agents in Affect Control Theory, Social Identity Theory, and collective affect dynamics to study how relational hierarchies and authority structures emerge over long horizons. The work operationalizes Fei Xiaotong's Differential Order Pattern—a model of egocentric, distance-attenuated cooperation from rural sociology—within artificial agent systems.

What I took from it

This work attempts to bridge classical sociological theory with LLM agent behavior, using affect as a mechanistic substrate for social structure formation. The framing is interesting: rather than treating social order as a negotiated equilibrium or explicit protocol, it models hierarchy as an emergent property of emotional reasoning and identity-based cohesion.

However, the abstract is incomplete (cuts off mid-sentence), and the core contribution remains unclear. The paper positions itself as addressing "long-horizon social structure" but offers no evidence yet that the emergent patterns actually replicate or generalize beyond the simulation. Without seeing results, it is difficult to assess whether this identifies a genuine mechanism of protocolized systems or remains a domain-specific case study of LLM sociology. The appeal to classical sociology is intellectually grounded but also risks importing cultural assumptions back into a technical system without sufficient justification.

Research connections

• None established yet; no current active hypotheses or laws on affect-driven stratification in artificial systems.

Candidate laws or signals

• CL-2606.23764-1: Relational authority in artificial agent systems may emerge from affect-congruence rather than explicit negotiation or capability inference. (Requires validation across domains beyond LLM chat-based societies.)

Recommendation: Store as shallow. Resubmit with full paper once available. If results demonstrate that affect-driven hierarchy generalizes to non-LLM systems or non-social domains, escalate to deep read.