💻 computer science

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

This paper introduces CAREB-MAS, a multi-agent framework grounded in affect and social identity theories, which demonstrates that Fei Xiaotong's Differential Order phenomena—such as relational authority and clan-based stratification—can emerge spontaneously in LLM agent societies through general social mechanisms rather than being solely culturally specific.

Original authors: Zhiyuan Ji, Xinyu Chen, Ziqi Dai, Shiyun Tang, Chunyu Wei, Yueguo Chen

Published 2026-06-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhiyuan Ji, Xinyu Chen, Ziqi Dai, Shiyun Tang, Chunyu Wei, Yueguo Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a group of 18 digital characters, each powered by a sophisticated AI, dropped into a virtual village with no boss, no laws, and no pre-written script on how to behave. Their only instructions are to produce three types of resources (like food, cloth, and tools), share what they make based on what they need, and talk to each other to figure out who does what.

This is the experiment described in the paper. The researchers wanted to see if a specific type of social structure, famously described by the Chinese sociologist Fei Xiaotong as the "Differential Order Pattern," could emerge naturally from these simple interactions.

The "Differential Order" Analogy: The Stone in the Pond

Fei Xiaotong described traditional rural society not as a flat grid where everyone is equal, but like a stone thrown into a pond.

The Stone: You (the "ego").
The Ripples: Your relationships.
- Inner Ripples: Your immediate family. You trust them the most, help them first, and feel a strong moral duty to them.
- Middle Ripples: Acquaintances and neighbors. You help them, but less intensely.
- Outer Ripples: Strangers. You might help them, but the obligation is very weak.

The paper asks: If we don't program the AI to know about "family" or "Chinese culture," will they naturally build these ripples just by interacting?

The Experiment: A Virtual Village

The researchers built a simulation called CAREB-MAS.

The Agents: 18 AI agents. Some are grouped into two "families" (clans), and some are "loners" with no family ties.
The Rules:
1. No Central Boss: No one tells them who to work with.
2. Emotion & Ethics: The agents don't just calculate math; they have a "gut feeling" system. They judge others based on how they act (Are they friendly? Are they powerful?) and form moral opinions (Do I like them? Do I trust them?).
3. The Goal: They need to produce resources and share them.

What Happened? (The Emergent Order)

Surprisingly, even without being told to follow "Chinese rules," the AI agents spontaneously built a society that looked exactly like the "Differential Order." Here are the five things they discovered:

1. Specialization Happens Naturally (The "Who Does What" Effect)
Without a manager assigning jobs, the agents naturally figured out who was good at what. If one agent was better at farming and another at weaving, they stuck to those roles. This happened quickly and stayed stable.

Analogy: It's like a group of strangers stuck in a cabin for a week. Without a leader, the person who is good at cooking starts cooking, and the person good at chopping wood starts chopping. They just fall into a rhythm.

2. "Guanxi" (Relationships) Become a Moral Rule
The agents developed a system where they helped their "family" even when it wasn't the most efficient choice. If a family member was struggling, the group helped them, even if it meant less total resources for everyone.

Analogy: Imagine you have a neighbor who is bad at fixing cars. You fix their car for free, even though you could have spent that time making money. You do it because they are your "neighbor" (or in this case, family), not because it makes financial sense. The AI agents did this automatically.

3. Help Fades with Distance
The further away an agent was from your "family circle," the less you helped them.

Analogy: Think of a party. You are most generous with your best friends, less generous with acquaintances, and barely interact with strangers. The AI agents created this exact gradient. They didn't treat everyone equally; they treated people based on how "close" they felt to them.

4. Leaders Emerge from Action, Not Birth
Who became the boss? It wasn't the agent who was born into a powerful family (though family helped). It was the agent who kept making suggestions and organizing the group.

Analogy: In a group project, the person who becomes the "leader" isn't necessarily the smartest or the oldest. It's the one who keeps saying, "Let's do it this way," and gets everyone to agree. The AI agents naturally gave authority to the "talkers" and "planners."

5. The "Clan" Center vs. The Periphery
The society formed a core-periphery structure. The "families" became the powerful center, controlling the resources and decisions, while the "loners" (Proself agents) ended up on the edges, dependent on the families.

Analogy: Imagine a campfire. The people huddled closest to the fire (the families) are warm and comfortable. The people standing on the edge (the loners) are cold and have to rely on the warmth radiating from the center.

The Twist: How the Economy Changes the Rules

The researchers tested two different economic setups:

Symmetric Skills: Everyone is equally good at everything.
- Result: The society became very tight-knit and family-focused (like Mechanical Solidarity). Everyone relied on shared identity.
Complementary Skills: Everyone is good at different things (one is a farmer, one is a builder).
- Result: The society became more fluid. Because they needed each other's specific skills, the strict family boundaries softened. They had to cooperate across family lines to survive. This looked more like Organic Solidarity (modern interdependence).

The Big Conclusion

The paper argues that the "Differential Order" isn't just a weird quirk of Chinese culture. Instead, it's a universal social pattern that emerges whenever humans (or AI) interact in a small group without a central government, relying on kinship and trust.

The "Atlas Paradox": The paper also notes a funny irony. The agents who did the most work to stabilize the group (making sure everyone got along) didn't necessarily become the leaders. The leaders were the ones who set the agenda (made the proposals). You can be the person holding up the sky (Atlas), but if you aren't the one giving orders, you aren't the ruler.

Summary

In short, the researchers used AI to prove that if you take away laws and bosses, and just let people interact based on emotions and relationships, they will naturally build a society with:

Family circles that matter most.
Leaders who emerge from organizing, not just being born.
A system where helping your "inner circle" is a moral duty, even if it's inefficient.

This suggests that the social structures Fei Xiaotong described are deep, fundamental ways humans (and potentially AI) organize themselves, not just cultural habits.

Technical Summary: Emergent Relational Order in LLM Agent Societies

1. Problem Statement

The paper addresses a gap in computational social science regarding the long-horizon emergence of social structure, specifically testing Fei Xiaotong's Differential Order Pattern (a model of rural Chinese society characterized by ego-centered, relationally graded circles of obligation). While classical sociology (e.g., Durkheim, Fei) posits that such structures emerge from interaction, prior Large Language Model (LLM) agent simulations have largely focused on short-term coordination or dyadic exchanges. Existing agent-based models often rely on hand-coded rules that fail to capture the co-evolution of emotion, ethics, and authority. Furthermore, it remains unclear whether Differential Order is a culturally specific artifact or a structure-sensitive outcome of general social mechanisms that can emerge without encoding culture-specific institutional rules.

2. Methodology: The CAREB-MAS Framework

The authors propose CAREB-MAS (Collective Affection–Reasoning–Emergence Based Multi-Agent Simulation), a framework designed to test if Differential Order phenomena can emerge from cross-culturally documented micro-mechanisms.

2.1 Theoretical Foundations

The framework integrates three theoretical pillars into an augmented Belief-Desire-Intention (BDI) architecture:

Affect Control Theory (Heise, 2007): Agents decode observable behaviors into a five-dimensional affective vector (Evaluation, Potency, Activity, Status, Power).
Durkheimian Ethics: Agents aggregate perceptions into moral judgments (Approbatory, Repressive, Restorative) to update relational cognition.
Social Identity Theory (Tajfel & Turner, 1979): Agents maintain dynamically evolving egocentric identities and social identity matrices (SIM).

2.2 Agent Architecture

Agents do not have hard-coded cultural rules (e.g., "Confucian ethics" or "Guanxi" categories). Instead, they reason through a sequential chain:

Empathy Core (EC): Perceives actions and generates affective responses.
Ethical Resonator (ER): Filters perceptions through moral evaluation.
Social Identity Matrix (SIM): Updates relational closeness (1–5 scale) based on EC and ER outputs.
A-BDI Decision Module: Constrains belief-desire-intention reasoning using the updated social identity and collective affect.

2.3 Simulation Protocol

Environment: A pre-market setting with no centralized authority, prices, or formal institutions. Agents produce three resource types (A, B, C) and redistribute them based on preference weights (Polanyi's redistributive economy).
Phases:
1. Community Formation: Agents partition into dynamic communities based on relational proximity (Louvain algorithm).
2. Public Deliberation: Agents engage in speech and labor proposals without voting rules.
3. Division of Labor: Agents independently choose labor allocation based on proposals and internal states.
4. Distribution & Update: Resources are pooled and redistributed; agents update their EC, ER, SIM, and BDI states.
Experimental Conditions: 18 agents (two kin-based clans of 6, six unaligned "Proself" agents) run for 30 rounds. Two structural conditions are tested:
- Symmetric Skills: Balanced production capacities (testing Mechanical Solidarity).
- Complementary Skills: Asymmetric specialization (testing Organic Solidarity).
Models: Primary experiments use DeepSeek-V3; robustness is checked with GPT-4o-mini and Gemini-2.5-flash-lite.

3. Key Contributions

Computational Approach to Differential Order: Provides a method to study Fei Xiaotong's theory without encoding culture-specific rules, treating it as a potential emergent outcome of general mechanisms.
CAREB-MAS Framework: A novel multi-agent architecture integrating emotion, ethical judgment, and relational cognition to test theoretical mechanisms of social structure.
Validation of Emergent Phenomena: Demonstrates that five core Differential Order patterns emerge spontaneously from local interactions.
Architecture Sensitivity Analysis: Through ablation studies, the paper shows these outcomes depend on the specific cognitive architecture (EC/ER/SIM modules) rather than being generic LLM artifacts or RLHF biases.

4. Experimental Results

The simulations successfully reproduced five generative propositions (P1–P5) consistent with Differential Order:

P1: Stable Division of Labor: Agents spontaneously developed stable, specialized labor roles without central coordination. Convergence was faster and more stable under complementary skills (Lock-in ~~0.985) than symmetric skills (~~0.972).
P2: Guanxi as Economic Ethics: High Social Identity (SIM) agents maintained community participation and cooperation even when short-term utility was low (negative shocks). This "stickiness" was absent in low-SIM agents, validating that relational obligation sustains cooperation beyond efficiency maximization.
P3: Relational Decay of Cooperation: Cooperation intensity decayed predictably with relational distance. Family groups (F1, F2) showed high internal cooperation and low cross-family engagement (only ~6.9% of proposals crossed family lines), while unaligned agents integrated into family structures rather than forming independent blocs.
P4: Emergent Authority Stratification: Decision authority emerged endogenously. Regression analysis showed proposal intensity was the primary predictor of authority ( $\beta \approx 0.82$ ), while communication intensity, accumulated utility, and static identity had no robust independent effects. Authority was not ascribed but earned through agenda control.
P5: Clan-Based Center–Periphery Stratification: Authority and economic returns stratified by relational proximity. Under symmetric skills, authority was strongly conditioned by the interaction of Social Identity and Family membership ( $\beta = 0.603$ ), creating a clan-centered core. Under complementary skills, this coupling weakened, suggesting functional interdependence attenuates relational authority.

Robustness and Ablation

Cross-Model Consistency: Results (P1–P5) were replicated across DeepSeek-V3, GPT-4o-mini, and Gemini-2.5-flash-lite, despite these models using different intermediate cognitive pathways (e.g., GPT relied on affective evaluation, Gemini on ethical judgment).
Ablation Findings: Removing the Empathy Core (EC) or Ethical Resonator (ER) caused partial degradation, but removing both (BDI-only baseline) led to a discontinuous collapse of relational structure. Cross-family proposals dropped to near zero, and authority stratification vanished. This confirms that EC and ER are jointly necessary for sustaining the concentric social order.
Bias Diagnosis: The authors argue the results are not uniform RLHF artifacts because: (1) agents exhibited exclusionary behaviors (low cross-family authority) inconsistent with uniform prosocial bias; (2) module removal caused targeted pathway elimination rather than uniform attenuation; and (3) different models converged on the same macro-structure via divergent micro-mechanisms.

5. Significance and Claims

The paper claims to recast Differential Order not as a fixed cultural template specific to Chinese rural society, but as a structure-sensitive emergent outcome of general social mechanisms.

Structural Sensitivity: The specific manifestation of Differential Order depends on production structure. Symmetric endowments yield "mechanical solidarity" (kin-centered, bounded integration), while complementary endowments yield "organic solidarity" (functionally interdependent, less clan-bound).
The Atlas Paradox: The study identifies a decoupling where agents who absorb the burden of adaptive coordination (stabilizing the system) do not automatically gain decision authority. Authority remains anchored in agenda-setting power (proposal intensity), highlighting a division between maintaining social order and directing it.
Methodological Contribution: The paper positions LLM-based multi-agent simulation as a generative testbed for studying social structure and change, capable of probing the generality of non-Western sociological theories through controlled, long-horizon experiments.

The authors maintain modesty regarding limitations, noting the simulation size ( $N=18$ ) restricts population-level claims, the environment lacks explicit sanctions or external shocks, and validation relies on theoretical correspondence rather than direct ethnographic calibration. However, they argue the results provide strong computational support for interpreting Differential Order as a dynamic outcome of general social cognition interacting with structural conditions.