Skip to content

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

🕒 Published (v1): 2026-06-22 14:25 UTC · Source: Arxiv · Venue: ACL 2026 · link

Why this paper was selected

ACL 2026; emergent authority stratification in LLM agent societies informs multi-agent coordination design

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CAREB-MAS is a multi-agent simulation harness that tests whether Fei Xiaotong's Differential Order Pattern—ego-centered, relationally graded social structure—emerges from general social mechanisms encoded in LLM agents, without any culture-specific rules. Eighteen agents running a four-phase loop across 30 rounds spontaneously reproduce five core sociological phenomena (labor specialization, guanxi ethics, cooperative decay, authority stratification, clan center-periphery). Results are architecture-dependent: the full EC→ER→SIM→A-BDI chain is necessary; BDI-only collapses P3–P5.

Problem

Prior LLM-agent simulations study short-horizon coordination or dyadic exchange and rely either on hand-coded cultural rules or generic BDI reasoning, making it impossible to determine whether non-Western sociological phenomena (Differential Order) reflect genuine general mechanisms or culture-specific assumptions. Conventional agent-based models (Sugarscape, Axelrod) cannot co-represent emotion, ethics, and authority dynamics simultaneously.

Method

CAREB-MAS (Collective Affection–Reasoning–Emergence Based Multi-Agent Simulation) chains three sociological theories into a sequential cognitive pipeline per agent per round:

  1. Empathy Core (EC) — Affect Control Theory: decodes observed behavior into a five-dimensional affective vector \((evaluation, potency, activity, status\text{-}conferral, power\text{-}assertion)\).
  2. Ethical Resonator (ER) — Durkheimian moral judgment: aggregates perceptions into approbatory/repressive/restorative attitudes, updating relational weights.
  3. Social Identity Matrix (SIM) — Social Identity Theory: maintains and updates egocentric relational distances \(d_{i \to j} \in \{1,\ldots,5\}\).
  4. A-BDI Decision Module — BDI extended with affective constraints from SIM; selects labor allocation \(\ell_i = (\ell_i^A, \ell_i^B, \ell_i^C)\) with \(\sum_r \ell_i^r = 1\).

Each round follows a four-phase loop: - Phase I — community formation via Louvain detection on mean relational proximity \(\bar{d}_{ij}\); - Phase II — \(R=2\) rounds of public deliberation (speeches + proposals, no enforcement); - Phase III — independent labor allocation and ex-post measurement of Decision Authority (DA) = proposal adoption frequency; - Phase IV — pooled redistribution by preference weights \(x_{ir} = \frac{(\sum_k o_{kr}) \cdot p_{ir}}{\sum_j p_{jr}}\), then EC→ER→SIM→A-BDI update.

Setup: 18 agents, 2 kinship clans (F1, F2; 6 members each) + 6 unaligned agents, 3 resource types, Leontief utility, 30 rounds, 5 random seeds. Primary model: DeepSeek-V3 (temperature 0.7); robustness on GPT-4o-mini and Gemini-2.5-flash-lite.

Key Contributions

  • Computational operationalization of Fei Xiaotong's Differential Order Pattern without encoding culture-specific behavioral rules.
  • CAREB-MAS framework integrating Affect Control Theory, Durkheimian ethics, and Social Identity Theory into an augmented BDI architecture.
  • Empirical validation of five generative propositions (P1–P5) with quantitative metrics (lock-in score, community match, DA regression, SIM×Family interaction).
  • Controlled ablations (EC-only removal, ER-only removal, BDI-only) across three LLMs demonstrating that P3–P5 require the full architecture.
  • Evidence that the same structural conditions yield qualitatively different emergent orders under symmetric (mechanical solidarity) vs. complementary (organic solidarity) production endowments.

Results

  • P1 (Division of Labor): Lock-in reaches 0.972 (symmetric) and 0.985 (complementary); per-round improvement → 0, confirming stable equilibrium.
  • P2 (Guanxi Ethics): High-SIM agents maintain community participation under negative utility shocks; low-SIM ties show sharp withdrawal; asymmetry stronger under complementary skills.
  • P3 (Relational Decay): Within-family proposals 34.2%, family→proself 34.3%, cross-family only 6.9% (symmetric); Proself agents lack self-boundary and absorb into family structures.
  • P4 (Authority Stratification): OLS: proposal intensity \(\beta = 0.820^{***}\) (symmetric), \(\beta = 0.780^{***}\) (complementary); communication intensity, cumulative utility, SIM show no robust independent effects.
  • P5 (Center–Periphery): SIM×Family interaction \(\beta = 0.603^{***}\) (symmetric); interaction disappears under complementary skills, where functional differentiation attenuates relational authority.
  • Ablations: BDI-only causes cross-family proposals to collapse (10.3% → 1.3%, Gemini) and DA to fall −39% (Gemini) / −70% (DeepSeek/GPT); single-module removal yields modest degradation; complete pipeline removal triggers discontinuous collapse of P3–P5.

Limitations

  • Simulates only 18 agents over 30 rounds; scaling to larger populations or longer horizons is untested.
  • Leontief utility and preference-weighted redistribution are deliberate simplifications of real economic mechanisms; price dynamics and markets are absent by design.
  • Decision Authority operationalized as proposal adoption frequency—a proxy that cannot distinguish charismatic authority from strategic repetition.
  • Communication Intensity is the only metric using an external LLM-as-Judge; its inter-rater reliability is not systematically validated.
  • Architecture behavior is LLM-dependent (results shown for 3 models); generalizability to other frontier models is not established.
  • The unaligned ("Proself") group lacks kinship ties by construction, confounding absence-of-kin with other individual-difference factors.

Relevance to Harnesses / Meta-Harnesses

CAREB-MAS is a concrete example of a theory-grounded agent harness: it composes multiple cognitive sub-modules (EC, ER, SIM, A-BDI) into a fixed sequential pipeline that governs every agent's reasoning loop, making the harness architecture itself the experimental variable (ablations remove modules, not agents). The four-phase social loop—community formation → deliberation → labor allocation → distribution+update—is a reusable orchestration scaffold that controls agent interaction topology, data flow, and evaluation independently of the LLM backbone, demonstrating how harness design choices (pipeline order, phase boundaries, enforcement vs. emergence) directly determine which macro-level phenomena can emerge. For harness researchers, the key insight is that the pipeline sequencing of perceptual, normative, and decision modules is load-bearing: removing any single stage degrades different emergent properties, showing that meta-harness designs must reason about module interaction order, not just module inclusion. The cross-model ablation methodology (same harness, swapped backbone) is also a transferable evaluation pattern for harness robustness testing.